DATA 2026 Abstracts


Area 1 - Data Analytics & Visualization

Full Papers
Paper Nr: 125
Title:

County-Level Mobility Patterns During the COVID-19 Pandemic in Virginia

Authors:

Rory Gardner, Savannah Drake and Behnaz Moradi-Jamei

Abstract: This paper analyzes county-level mobility shifts across Virginia using Google Community Mobility Reports from 2020–2022. Mobility trends were examined across five movement categories (Retail and Recreation, Grocery and Pharmacy, Transit Stations, Workplaces, and Residential), measured as percent change relative to a January–February 2020 baseline. After filtering counties with insufficient data, remaining gaps were imputed using bidirectional linear interpolation with within-county forward- and backward-fill. Counties were classified as metropolitan or non-metropolitan, and principal component analysis identified a “mobility away from home” component (PC1) explaining 66.18% of total variance. Exact county-level permutation inference showed a statistically significant difference in mean PC1 scores between metropolitan and non-metropolitan counties (p = 0.0022), with a large effect size (Cliff’s δ = 0.7655). This contrast remained strong after residualizing statewide year-month effects. Category-specific residualized comparisons showed that Grocery and Pharmacy and Workplaces contributed most strongly to the metropolitan versus non-metropolitan difference. Shock analyses showed a broader contrast than workplaces alone, with the clearest adjusted evidence in Grocery and Pharmacy. A formal cross-category test showed that workplace mobility was among the deepest declines, though not significantly deeper than transit. Overall, metropolitan counties experienced sharper, more persistent reductions in out-of-home mobility, highlighting regional differences masked by statewide averages.
Download

Paper Nr: 185
Title:

The EDIAQI Dashboard: Data Monitoring and Troubleshooting for Cross-European Indoor Air Quality Comparison

Authors:

Daniela Dejaco, Han Tran and Heimo Gursch

Abstract: The EDIAQI Dashboard is a combined tool that is capable of doing an extensive data analysis and data quality troubleshooting for complex data scenarios in Indoor Air Quality monitoring. The Dashboard offers the possibility to compare assessments of Indoor Air Quality across Europe and serves as a consistent data visualization tool for the EDIAQI project. The objective of the data analysis and troubleshooting is not to merely postprocess the data, but to identify and solve data issues directly at the measuring unit. The corrected data with sufficient quality can be used in the same tool for a Cross-European Indoor Air Quality Comparison, revealing patterns, analogies and differences in the various institutions and private households for which indoor air parameters were recorded. The EDIAQI Dashboard represented an important milestone for the project, since it was capable of bridging the gap between the very distinct parts of the project.
Download

Short Papers
Paper Nr: 34
Title:

Tracking Urban Atmospheric Pollutants Using Sentinel-5P Satellite Data

Authors:

Alice Gomez-Cantos and Henry O. Velesaca

Abstract: Urban nitrogen dioxide (NO2) is a key indicator of combustion-related air pollution and exhibits strong spatial and temporal variability in cities. This study presents a satellite-based framework for tracking urban NO2 pollution using tropospheric column observations from Sentinel-5P/TROPOMI over Guayas Province, Ecuador. Rather than estimating surface concentrations, the methodology emphasizes robust distributional metrics, including the median and upper-tail percentiles (P90, P95, and P99), to characterize background conditions and localized pollution extremes at the canton scale. Multi-year satellite observations are aggregated annually and analyzed using unsupervised K-means clustering to identify characteristic pollution regimes without predefined thresholds. Results show that highly urbanized cantons consistently exhibit elevated extreme NO2 values and greater variability, while less urbanized areas display lower and more homogeneous patterns. The proposed approach provides an interpretable and scalable tool for urban air-quality assessment in data-scarce regions using satellite observations alone. The implementation is publicly available on GitHub https://hvelesaca.github.io/sentinel-5P-clustering/.
Download

Paper Nr: 166
Title:

A Context-Aware Framework for Environmental Resilience in Maritime Operations

Authors:

Nadeem Iftikhar

Abstract: Vessel performance is influenced by both internal machinery and external operating conditions such as weather, sea state, currents, and route context. Therefore, a drop in efficiency does not always indicate technical degradation. This paper presents an interpretable framework for analyzing environmental resilience in maritime operations, where environmental resilience refers to the ability of a vessel to preserve performance under increasing environmental stress. The framework combines vessel telemetry, environmental information, and route context to assess four aspects of operation: efficiency after accounting for environmental effects, sensitivity to environmental stress, route difficulty, and operational stability. The framework is evaluated using representative operating periods, synthetic stress tests, and illustrative regional summaries. The results show that the method helps distinguish environmentally driven performance loss from performance changes that are less well explained by external conditions, while supporting context-aware comparison across operating conditions.
Download

Paper Nr: 215
Title:

Intelligent Analysis with Visualisation of Health Data

Authors:

Inês Ribeiro, Alexandra Oliveira and Brígida Mónica Faria

Abstract: Intelligent data analysis and visualisation transform large volumes of information into practical knowledge, enabling faster and more informed decision-making. Visual tools such as charts and interactive dashboards simplify the interpretation of complex data, facilitating pattern recognition and trend identification. The growing digitalisation of healthcare has generated vast datasets with significant analytical potential. This study aims to support the analysis of large healthcare datasets and improve clinical decision-making through structured data visualisation. A taxonomy of visualisations was developed following a systematic approach to ensure clarity, comprehensiveness, and applicability. Existing literature and classification frameworks were reviewed to identify key criteria, and a total of 59 visualisations were organised according to their purpose, analytical objective, and variable type. The taxonomy was validated using real-world data and implemented in an interactive Power BI environment, enabling dynamic exploration of categories and examples. Visualisations were applied to a public dataset of patients with chronic obstructive pulmonary disease, alongside the development of an interactive clinical profile card summarising key patient information. A convenience sample of 40 healthcare professionals evaluated the outputs. Simpler visualisations were perceived as more accurate and useful, and 95% of participants preferred interactive over static formats.
Download

Paper Nr: 45
Title:

Compositional Data Analysis for Quality Assessment and Root Cause Analysis in Manufacturing

Authors:

Sonja Strasser, Herbert Jodlbauer and Sezgin Kutlu

Abstract: Compositional data (CoDa), such as product recipes and material mixtures, are ubiquitous in manufacturing but present analytical challenges due to their constant-sum constraint and non-Euclidean geometry. Effective root cause analysis in complex compositional datasets requires methods that can filter relevant cases and provide intuitive visualizations. This paper presents a systematic framework for preparing, analyzing, and visualizing CoDa to support quality assessment and root cause identification in manufacturing. The framework begins with an exploratory phase that introduces metrics to quantify quality-class separability across large product portfolios. A subsequent diagnostic phase employs centroid ratios and visualizations, including parallel coordinate plots and Principal Component Analysis, to identify specific material deviations. For complex scenarios with overlapping quality classes, clustering is applied to uncover latent production regimes and failure mechanisms. The framework’s application in a case study demonstrates its efficiency and interpretability for quality investigations, supporting data-driven decision-making for process managers.
Download

Paper Nr: 232
Title:

Dialysis Irregularity in India: The Role of Geographic Access and Patient-Reported Travel Burden

Authors:

Sachin Bodke, Narayan Rangaraj and Vishwanath Billa

Abstract: Access to maintenance haemodialysis is often assessed using geographic distance to the dialysis centre, although distance alone may not fully reflect the transportation burden experienced by patients. We analysed registry data from 21,569 adult patients receiving in-centre haemodialysis across 256 centres in 14 Indian states, along with a travel survey involving 2,087 patients. Two outcomes were evaluated: frequency-based dialysis irregularity, defined as receiving ≤2.5 dialysis sessions per week, and structural irregularity, defined as patient-reported missed, delayed, or shortened dialysis sessions due to transportation-related barriers. Multivariable logistic regression models with centre-clustered standard errors were used. In the registry cohort, older age and greater travel distance were associated with higher odds of dialysis irregularity, while treatment at public and large urban centres was associated with lower odds. In the survey cohort, frequency-based irregularity was more strongly associated with treatment at hospital-based centres (OR 3.94, 95% CI 2.01–7.71), longer travel time (OR 1.09, 95% CI 1.02–1.17), and caregiver dependence (OR 1.41, 95% CI 1.07–1.86). In contrast, structural irregularity was primarily associated with day-to-day transportation burden, particularly longer travel time (OR 1.11, 95% CI 1.04–1.17) and higher travel cost (OR 1.07, 95% CI 1.00–1.14). Frequency-based irregularity was associated with centre type, age, caregiver dependence, and travel time; structural irregularity was associated with travel time and travel cost. As an observational study, these results describe associations rather than causal effects.
Download

Area 2 - Data Engineering, Infrastructure and Business Applications

Full Papers
Paper Nr: 122
Title:

Data-Centric Lifecycle Integrity Validation for IEC 61508 Using Executable Documentation

Authors:

Padma Iyenghar, Christopher Zimmerman and Claudio Gregorio

Abstract: Industrial cyber-physical systems rely on structured safety lifecycle in which hazards, risks, requirements, architecture, and verification activities are related across defined phases. In practice, these elements are distributed across heterogeneous documents, making structural consistency and traceability difficult to maintain. This paper presents an approach in which lifecycle elements are encoded as structured documentation artifacts with explicit types and relations. From a data-centric perspective, these artifacts are treated as structured data objects whose relations form a constrained graph representation of lifecycle information. A controlled vocabulary and constrained relations define a consistent representation of lifecycle data. The artifacts are processed through a documentation build workflow that acts as a deterministic data processing and validation pipeline, automatically generating traceability views and completeness checks. Structural constraints over artifact relations cover verification, evidence linkage, SIL allocation, and hazard coverage. The approach is demonstrated on a reproducible IEC~61508 SIL~2 Emergency Stop example, with an open demo site and repository that provide all sources and build scripts. Violations introduced by removing required relations are detected as non-empty coverage tables. The results show that lifecycle integrity can be evaluated as a structural property of the artifact graph, enabling early, machine-checkable detection of missing relations.
Download

Paper Nr: 131
Title:

Decision Support System for the Use of Robotic and Intelligent Process Automation in Small and Medium-Sized Enterprises

Authors:

Ann-Kathrin Holaj and Sascha Alpers

Abstract: Small and medium-sized enterprises (SMEs) face growing pressure to automate their business processes but often lack the resources to evaluate automation initiatives systematically. Robotic Process Automation (RPA) addresses rule-based, structured processes, whereas Intelligent Process Automation (IPA) combines RPA with artificial intelligence (AI) to handle more complex tasks. However, a systematic framework that helps SMEs decide whether a process is suitable for RPA, IPA, or neither is still missing. This paper addresses that gap by presenting a prototype decision support system (DSS) called “Automation Fit”, developed following a design science research approach. Based on a literature review, 32 criteria across multiple dimensions were identified, from which a questionnaire and a scoring model were derived. Expert validation through eight semi-structured interviews provided preliminary support for the relevance and practical feasibility of the assessment approach. The resulting prototype was evaluated in two qualitative interviews focusing on user experience and practical applicability in an SME context. The findings suggest that the prototype can support the initial stages of automation assessments, while identifying areas for improvement relating to interface terminology and the clarity of technically complex questions.
Download

Paper Nr: 160
Title:

Beyond Context Windows: Data Transformations at Scale with LLMs

Authors:

Abhigith Abraham, Fariz Rahman, Fadil Rahman, Monika Janchevska and Tanaya Amar

Abstract: Applying large language models (LLMs) to real-world tabular datasets at scale is subject to a fundamental mismatch between model context limits and enterprise data size. Existing approaches address this by truncating data or fitting everything into a single prompt - methods that do not scale and sacrifice row-level semantic understanding. We present a framework for scalable, dataset-scale LLM-based data transformation. It processes data row-by-row in batches over lazily evaluated, distributed dataframes, enabling LLM-based transformations on arbitrarily large datasets without loading them into memory. Three composable primitives cover the most common patterns: MAP for field-level enrichment, FILTER for row selection, and REDUCE for semantic deduplication via embedding-based similarity search. An agentic orchestration layer accepts a plain-language goal, plans the required primitive sequence, and routes operations that require no semantic reasoning to generated Python code. We evaluate the framework on five domain-specific benchmarks. On a 51,291-row agentic enrichment task, the pipeline achieves 96.2% accuracy in 9.4 minutes. On a 3-million-row deduplication task, REDUCE completes in 1.90 hours. Single-primitive benchmarks reach 99.6% coverage on category extraction, Macro F1 of 0.97 on sentiment classification, and 96% accuracy on compliance filtering. The framework is released as open source at https://github.com/vitalops/datatune.
Download

Short Papers
Paper Nr: 49
Title:

Graph-Based Named Entity Management System with Ontology-Supported Matching Capabilities

Authors:

Andrea Leoni, Andrea Molinari and Simone Sandri

Abstract: Entity Matching (EM) is a core challenge in data integration, requiring the identification of records that refer to the same real-world entity across heterogeneous sources. Practical experience shows that overall performance depends on end-to-end system design rather than isolated algorithms: candidate generation, threshold calibration, provenance tracking, consolidation policies, and iterative error analysis often determine effectiveness. Ontologies, knowledge graphs, and persistent identifiers provide semantic context and stable references, but introduce additional complexity in handling uncertainty and evolving representations. We present a Named Entity Management System (NEMS) that integrates entity lifecycle management with scalable matching in a unified workflow for knowledge graph creation. Instead of treating reconciliation as post-processing, NEMS embeds matching and validation during ingestion, combining attribute-level similarity with graph-structured and ontology-aware signals to guide merge decisions. By integrating canonical identifiers, provenance tracking, and configurable decision thresholds, NEMS enables conservative merging, incremental updates, and explainable outcomes. The architecture accommodates diverse matching paradigms while leveraging structural context, providing a robust foundation for scalable and consistent entity integration.
Download

Paper Nr: 67
Title:

Architectural Alternatives for Data Distribution in Autonomous Driving: From Real-Time Performance to Semantic Interoperability

Authors:

Mario Martín-Pérez and Marta Zorrilla

Abstract: The transition from Industry 4.0 and cloud-based paradigms to the IoT Continuum represents a fundamental shift in data processing models, moving from cloud-centric systems to distributed architectures with processing capabilities at the edge and IoT sensors. This paper evaluates two architectural alternatives for data transmission in autonomous driving: a performance-driven approach based on DDS-Kafka and a context-driven semantic approach based on FIWARE. While the former focuses on decentralized real-time communication and high-throughput streaming, the latter prioritizes semantic interoperability across IoT domains and context management. Through a qualitative analysis of their capabilities, we demonstrate how both architectures fulfill different requirements within the computing continuum, providing a roadmap for choosing the appropriate framework based on latency constraints, throughput or infrastructure integration needs.
Download

Paper Nr: 89
Title:

Closing the Loop: How Fleet IoT and Explainable AI Can Target Severe-Crash Risk on U.S. Roadways

Authors:

Suryakant Kaushik

Abstract: Getting crash severity predictions right matters-it shapes where agencies invest and which safety countermeasures they deploy. This paper presents a practical, data-driven framework using large-scale open datasets to understand how environmental and roadway factors influence crash outcomes across the United States from 2016 to 2023. Using the U.S. Accidents database (7.7M+ records), I standardized and analyzed variables including weather, visibility, precipitation, temperature, time, junction presence, and traffic-signal control. Two models were evaluated: a logistic regression baseline for interpretability and an XGBoost model to capture nonlinear interactions. On an 80/20 split, XGBoost achieved 0.70 accuracy (macro-F1: 0.61; ROC-AUC: 0.715) with minimal train-test divergence (0.001), indicating strong generalization. The precision-recall AUC of 0.374-over double the severe-class base rate (~0.17)-demonstrates reliable performance despite class imbalance. SHAP analysis identified wind speed, traffic signals, and weather conditions as key drivers: adverse weather and higher winds increased severity risk, while signalized intersections reduced it. Regionally, the Midwest and Southeast showed higher severity rates, with a decline after 2019. Beyond identifying crash-severity risk factors, this study contributes a reproducible national-scale modeling and explainability pipeline that links open-data fusion, severe-class prediction, SHAP-based interpretation, privacy-aware deployment, and validation-oriented safety interventions for roadway safety decision-making.
Download

Paper Nr: 132
Title:

A Holistic Framework for Assessing and Representing Vehicle Measurement (Time-Series) Data Quality via a Single Quality Score

Authors:

Sven Dominka and Henrik Smith

Abstract: Vehicle measurement data is a valuable basis for the development of simulation and machine learning models in the automotive industry. However, the quality of this data, particularly of the time-series measurement signals, is a critical factor that is often overlooked. This paper introduces a framework for assessing the quality of time-series signals. We propose a multi-dimensional approach based on three pillars: plausibility, coverage, and uniformity. For each pillar, we define quality metrics to provide a more nuanced evaluation. These metrics are then aggregated into a single, comprehensive quality score. A proof-of-concept demonstrates that our framework can successfully assess the quality of vehicle measurement data and hereby support the reuse of vehicle measurement files. The proposed approach facilitates a more reliable and trustworthy data selection process, enabling engineers to choose datasets that are genuinely suitable for demanding applications.
Download

Paper Nr: 179
Title:

FAIR Data Management from Collection to Exploitation: The RAISE Suite Project

Authors:

Evdokimos Konstantinidis, Gorka Epelde, Dimosthenis Natsos, Despoina Petsani, Anastasia Valtopoulou, Elli Papadopoulou, Mikel Hernandez, Dimitris Bamidis, Panagiotis Sarigiannidis, Andreas Symeonidis, Alexandros Chatzigeorgiou and Panagiotis Bamidis

Abstract: Despite growing mandates for FAIR (Findable, Accessible, Interoperable, Reusable) data practices, adoption remains limited by time constraints, insufficient incentives, and curation burden. Current infrastructures treat FAIR compliance as a retrospective activity, leaving a gap at the data collection stage. This paper introduces RAISE Suite, a DMP-centric, FAIR-by-design platform designed to support and automate data management from collection to dataset generation. By elevating the machine-actionable Data Management Plan (maDMP) to an orchestration backbone, RAISE Suite embeds metadata automation, provenance tracking, anonymization, and quality monitoring into scientific workflows. Integration with EOSC-RAISE provides a federated, controlled-access environment for data publication and algorithm-to-data processing. To be demonstrated across five pilots in health, mobility, forestry, oceanography, and food waste management, RAISE Suite aims to advance reproducible science through procedural FAIRness and reduced overhead.
Download

Paper Nr: 202
Title:

Hf2dcat: Bridging Hugging Face Metadata and DCAT-AP for Interoperable Data Catalogs

Authors:

Fangfang Wang

Abstract: Machine learning (ML) datasets and models are being shared at unprecedented scale: Hugging Face alone hosts over 2.8 million models and nearly one million datasets as of April 2026. At the same time, standards and metadata specifications such as RDF, DCAT, Croissant, and ML-specific schemas, ontologies, and application profiles provide a foundation for the interoperable representation, exchange, and integration of ML metadata. Yet these ecosystems remain largely disconnected: ML metadata stays platform-specific, while Semantic Web and catalog-oriented standards are rarely applied to real-world ML resources, limiting their discoverability, interoperability, reuse, and governance. This paper presents hf2dcat, a framework that bridges this gap by transforming Hugging Face metadata into DCAT-AP-compliant RDF records and enriching them with selected ML-specific semantics. Evaluation on the 100 most downloaded datasets and 100 most downloaded models shows that: (i) all accessible resources (100 datasets, 96 models) were converted to valid DCAT-AP RDF; (ii) Croissant metadata was preserved and exposed through linked metadata distributions for 95 of 100 datasets; (iii) processing scaled approximately linearly, requiring about 20 minutes for the full set; and (iv) the generated RDF records were successfully ingested into a local DCAT-AP-compatible Piveau catalog instance. By operationalizing DCAT-AP for catalog interoperability, Croissant for machine-readable dataset metadata and structure, and selected ML-domain vocabularies for semantic enrichment, hf2dcat provides a deployable bridge from platform-native ML metadata to FAIR-oriented open data infrastructures.
Download

Paper Nr: 225
Title:

Toward Uncertainty-Aware and Structure-Sensitive Repair of Missing Activity Labels in Process Mining Event Logs

Authors:

Lerina Aversano, Felice Franchini, Debora Montano and Chiara Verdone

Abstract: Missing activity labels are particularly harmful in process mining because they weaken the structural interpretation of a trace while leaving the trace otherwise observable. We therefore treat repair not as isolated token completion, but as process-aware reconstruction under structural and informational constraints. We formulate the task as selective, structure-sensitive event-log repair motivated by behavioral preservation. The current instantiation combines a bidirectional recurrent encoder and a self-attention encoder through late probabilistic fusion, while exposing posterior-score information that can support automatic repair, candidate-set repair, or abstention. We evaluate this pipeline on BPI Challenge 2017 under single-event, multi-event, and contiguous span masking, on both a controlled subset of 1,000 traces and the full log with 31,509 traces and 1,202,267 events. The deterministic late-fusion output reaches 84.8%, 91.5%, and 89.3% accuracy on subset single, multi, and span masking, and 99.3%, 99.1%, and 98.8% on the corresponding full-log settings. A strong XGBoost baseline confirms that the benchmark is highly amenable to attribute-aware prediction, so raw accuracy should be interpreted as only one part of the repair problem. The main contribution is therefore methodological rather than architectural. We argue that missing activity-label repair should be evaluated through uncertainty exposure, selective intervention, leakage resistance, and downstream structural diagnostics in addition to label-level success. The current experiments instantiate only part of that agenda; realistic missingness, process-constrained reranking, stricter leakage-resistant splits, richer downstream validation, repeated-seed evaluation, and explicit runtime-cost analysis remain important future work.
Download

Paper Nr: 158
Title:

Data Quality Validation in Enterprise Master Data Management

Authors:

Philipp Korom, Christine Dominka-Kiss, Anna-Christina Glock and Lisa Ehrlinger

Abstract: High data quality (DQ) is a core requirement in enterprise master data management (MDM), since poor data directly degrades downstream business decisions. Rule-based data validation is the predominant approach in MDM settings. However, deterministic rules are mainly good in detecting syntactic violations and often miss semantically invalid records. This paper therefore investigates an alternative approach by conducting a comparative evaluation of zero‑shot large language model (LLM)–based validation using Microsoft 365 Copilot and a rule‑based validation using the data management platform Ataccama ONE. The evaluation was carried out within the MDM program of the Austrian Post. In a first controlled experiment with 10,000 synthetic customer records, Copilot achieved an F1‑score of 1.00 for semantically complex error types, such as incorrect salutations and placeholder values, which are typically not captured by deterministic rules. A second experiment on 10,000 real Customer Relationship Management (CRM) records showed that both approaches offer distinct advantages, and we therefore recommend their combination. While the rule-based approach provides deterministic and consistently accurate results with high computational efficiency, it remains limited to predefined error patterns and therefore fails to detect semantic errors. By contrast, the LLM can identify such semantic inconsistencies and additionally suggests corrections.
Download

Paper Nr: 230
Title:

Data Acquisition, Refinement, Knowledge (Dark)

Authors:

Óscar Oliveira and Bruno Oliveira

Abstract: Dark data refers to the vast amount of information collected during regular business activities but largely unused for analytics or decision-making. The challenges of leveraging dark data for knowledge discovery are multifaceted, including its lack of visibility, unstructured formats, and fragmentation across disparate systems, which hinder access and analysis. Moreover, dark data often lacks proper documentation, consistent naming conventions, and well-defined relationships, complicating its interpretation and limiting its potential value. In this article, we present a structured approach to uncover business insights from underutilized databases by integrating large language models (LLMs) into the process. This approach is cost-effective, adaptable, and can be tailored to the specific needs of organizations, addressing key barriers such as the absence of metadata and data inconsistencies. This approach leads to a deeper organizational understanding of data and its latent potential. Future work will focus on refining this approach and extending it to unstructured data sources.
Download

Area 3 - Data in Industry and Emerging Trends in Data

Short Papers
Paper Nr: 163
Title:

An Open Data Innovation Platform for Sustainable Waste Reduction in Cable Manufacturing

Authors:

Bilel Elayeb, Michael Lecointre, Maher Jridi and Benoît Lardeux

Abstract: This paper proposes an Open Data Innovation Platform (ODIP) designed to support sustainable waste reduction through advanced data integration, storage, and analytics. Developed within the ACOME industrial environment, the platform leverages a data lake architecture to unify historical and real-time production data, enabling predictive modeling and decision support. A machine learning–based production line prediction simulator, built using a Random Forest algorithm and enhanced with SMOTE to address class imbalance, is introduced to identify optimal production configurations that minimize waste. Experimental results demonstrate a significant improvement in prediction performance, with balanced gains in precision, recall, and F1-score for both waste and non-waste classes. The platform further enables actionable recommendations for selecting optimal production lines for different cable categories, contributing to reduced material losses and improved operational efficiency. The proposed approach highlights the potential of open data platforms combined with artificial intelligence to transform waste management from a reactive to a proactive process in industrial contexts. It provides a scalable and transferable framework for integrating data-driven sustainability practices into manufacturing systems.
Download

Paper Nr: 195
Title:

When Retrieval Beats Generation: A Three-Condition Framework for AI-Driven Molecular Linker Design, with PROTAC as a Case Study

Authors:

Yeonju Jeong, Jae-Mun Choi and Young-Kuk Kim

Abstract: Generative AI has rapidly expanded into molecular linker design, with models such as SyntaLinker, DeLinker, DRlinker, and Link-INVENT being widely proposed. Yet these generation-first approaches often fail to translate into experimental validation in real medicinal chemistry workflows. We argue that this gap arises from a paradigm mismatch rather than implementation immaturity. Under three simultaneous conditions (data scarcity, multi-constraint satisfaction, and interpretability requirements), generation-first approaches face the following structural limitations: the synthesizability of generated molecules cannot be reliably guaranteed at the design stage, outputs are disconnected from medicinal chemists’ interpretive language, and sample complexity exceeds what available data can support. We take PROTAC linker design as a representative case where these three conditions simultaneously hold, and provide quantitative evidence of distributional mismatch between PROTAC linkers and general small-molecule linkers. As an alternative, we propose a property-profile-driven library retrieval approach in which physicochemical profiles are predicted based on the design context and candidates are selected from existing libraries accordingly. We further outline a hybrid research agenda integrating retrieval with generation for data-scarce molecular design.
Download

Paper Nr: 236
Title:

Threats of Retractions of Scientific Works to Information Ecosystem: How Can AI Help?

Authors:

Peiling Wang

Abstract: This study analyzed a dataset of 68,977 records from the Retraction Watch database to investigate the serious threats that retracted scientific publications pose to information ecosystems. The growing volume of retracted and corrected journal and conference papers implicates more than 178,000 named authors affiliated with institutions across over 180 countries and spanning multiple disciplines and subject areas. Both post-retraction and prior-to-retraction citations continue to influence researchers and information systems when flawed works are used as valid or legitimate sources. Recognizing the threats calls for more effective post-retraction interventions beyond the current approaches: notices of retracted papers by publishers and indexes of the retracted papers in databases. Taking into consideration human information behavior, this study proposes an AI-assisted model (AI-PRAM) to support both researchers and editorial workflows.
Download

Paper Nr: 212
Title:

AI-Enabled, Doctrine-Compliant Decision Superiority for CBRN and Hybrid Threat Environments

Authors:

Metin Çet, Sara Abri, Rayan Abri and Salih Cetin

Abstract: Contemporary CBRN threats increasingly emerge within hybrid, multi-domain environments characterised by ambiguity, distributed data streams, and compressed decision timelines. Existing NATO-compliant systems remain largely reactive, deterministic, and dependent on human-initiated processes, generating decision delays of 20–45 minutes-well beyond the critical 10–15-minute early-warning window required to prevent mass casualties. The study identifies five major operational capability gaps and proposes the AI-Enabled CBRN Decision Superiority System (AI-CDSS), a six-layer, doctrine-compliant architecture combining predictive analytics, high-resolution environmental modelling, civil-military data fusion, and explainable AI (XAI). Its core innovation is the trust-centred integration of advanced technologies within a framework that embeds NATO ATP-45 decision logic as formal finite-state machines, quantifies uncertainty through calibrated confidence intervals, and preserves human-in-the-loop override authority. Simulation-based projections demonstrate a 60–80% reduction in decision latency, a 66% decrease in hazard-zone area error, and removal of manual doctrinal translation steps.
Download

Area 4 - Data Privacy & Security, Ethics & Governance

Full Papers
Paper Nr: 193
Title:

A Serverless Client-Side Privacy Index for Sensitive Data Processing

Authors:

José Gameiro, José Luís Oliveira and João Rafael Almeida

Abstract: The widespread adoption of data-driven systems has intensified concerns regarding the protection of sensitive information and the assessment of privacy risks. Although numerous privacy-preserving techniques and models have been proposed, quantifying and interpreting the level of privacy achieved remains challenging, particularly for non-expert users. This paper introduces the Privacy Index, a unified metric designed to aggregate multiple privacy-related factors into a single, interpretable score. The proposed approach integrates the validation of established privacy models, detection of anonymization techniques, and estimation of re-identification risk under different attacker assumptions. To support practical usage, we develop a serverless, client-side web application that automatically processes structured datasets by classifying attributes and computing the Privacy Index without transmitting data externally. Experimental evaluation demonstrates that the attribute classification component achieves average confidence levels above 88% across multiple datasets, correctly identifying all direct identifiers with high reliability. The system successfully validates privacy models such as k-anonymity and l-diversity and effectively distinguishes between poorly and well-anonymized datasets. The tool is open-source, its code is available at https://github.com/ieeta-mith/DataPrivScore and a demo is available through https://ieeta-mith.github.io/DataPrivScore/.
Download

Paper Nr: 210
Title:

From Privacy-Utility Trade-Offs to Policies: Optimized Anonymization Recommendations for Data Trustees in Data Spaces

Authors:

Michael Steinert, Bekzod Nazarov, Thorsten Reitz and Daniel Tebernum

Abstract: As data spaces emerge to facilitate sovereign data exchange, geospatial data has become a critical resource across various domains. However, sharing sensitive geodata, such as cadastral property records, presents a privacy-utility trade-off. While the European Data Governance Act (DGA) establishes data trustees as the technical intermediaries authorized to perform necessary anonymization, a gap remains between theoretical spatial privacy algorithms and their practical, policy-driven application. This research addresses this gap by presenting an empirical evaluation of 11 spatial anonymization methods applied to complex polygon geometries. Using a dataset of 2,147 forested cadastral parcels, we quantify the trade-offs using privacy metrics (k-anonymity, differential privacy ε) and utility measures (Hausdorff distance, area deviation, centroid shift). Our results identify algorithmic failure modes, such as the “sparse forest” phenomenon, and demonstrate the improved performance of combined hybrid approaches. To operationalize these findings, we integrate our results into the architecture of data trustees operating within various domain-specific data spaces. We demonstrate how the empirically derived, use-case-specific anonymization guidelines can be translated into machine-interpretable Open Digital Rights Language (ODRL) policies. By mapping spatial transformation algorithms to ODRL constraints, this research provides data trustees with an approach to automatically enforce spatial privacy and sovereignty within data spaces.
Download

Paper Nr: 228
Title:

Metadata Management in Wind Energy Research Data: The Windlab Knowledge and Data Hub

Authors:

Miroslav Puskaric and Nina Buck

Abstract: The increasing scale and complexity of research data require efficient infrastructures for data management, interoperability, and long-term accessibility. Large collaborative research projects often produce heterogeneous datasets across multiple institutions, making standardized approaches to data management essential. Challenges include ensuring consistent data stewardship, developing comprehensive yet usable metadata model for diverse data, and integration into data analysis workflows. This paper presents the requirements analysis, system architecture, and metadata standardization framework implemented within the WindLab Knowledge and Data Hub. Functional and technical requirements were identified to ensure integration with distributed computing infrastructures and compliance with FAIR (Findable, Accessible, Interoperable, Reusable) data principles. To balance descriptive completeness with usability, a simplified metadata schema was developed, incorporating core elements from DataCite alongside domain-specific vocabularies. The implementation demonstrates how open-source data management platforms and standardized metadata frameworks are combined to support collaborative wind energy research and promote reproducible data-driven research while maintaining usability.
Download

Short Papers
Paper Nr: 68
Title:

Cryptography-Enhanced Data Spaces: Secure Cross-Domain Data Sharing via IDS Connectors

Authors:

Saneea Malik, Gabriella Laatikainen, Ilkka Niskanen, Juha-Pekka Soininen and Jukka Manner

Abstract: Data spaces enable controlled data sharing across organizations while preserving data sovereignty through policy-based governance mechanisms. Frameworks such as the International Data Spaces (IDS) provide standardized infrastructures for secure data exchange; however, they lack native mechanisms for performing computations on confidential data and for securely combining datasets across domains. In current implementations, datasets typically need to be decrypted before analysis, which limits the applicability of data spaces in privacy-sensitive environments and restricts the potential for cross-domain data fusion. To address this limitation, this paper presents an architecture implemented within the TRUSTEE platform that enables privacy-preserving computation on encrypted datasets hosted by IDS connectors. The proposed approach integrates homomorphic encryption (HE) workflows with standard data space connector infrastructures, allowing analytical operations to be performed directly on encrypted data while preserving existing governance and policy enforcement mechanisms. As a result, raw datasets remain protected within the provider environment throughout the computation lifecycle, and only the final computation results are shared with authorized participants. The proposed architecture is validated through a proof-of-concept demonstrating encrypted data fusion across multiple domains using TRUSTEE and IDS-based infrastructures. The evaluation confirms the feasibility, interoperability, and practical applicability of the approach for enabling secure secondary use of data and privacy-preserving cross-domain analytics in federated data space environments.
Download

Paper Nr: 108
Title:

ThreatCore: A Benchmark for Explicit and Implicit Threat Detection

Authors:

Davide Bruni, Carlo Bardazzi and Maurizio Tesconi

Abstract: Threat detection in Natural Language Processing lacks consistent definitions and standardized benchmarks, and is often conflated with broader phenomena such as toxicity, hate speech, or offensive language. In this work, we introduce ThreatCore, a public available benchmark dataset for fine-grained threat detection that distinguishes between explicit threats, implicit threats, and non-threats. The dataset is constructed by aggregating multiple publicly available resources and systematically re-annotating them under a unified operational definition of threat, revealing substantial inconsistencies across existing labels. To improve the coverage of underrepresented cases, particularly implicit threats, we further augment the dataset with synthetic examples, which are manually validated using the same annotation protocol adopted for the re-annotation of the public datasets, ensuring consistency across all data sources. We evaluate Perspective API, zero-shot classifiers, and recent language models on ThreatCore, showing that implicit threats remain substantially harder to detect than explicit ones. Our results also indicate that incorporating Semantic Role Labeling as an intermediate representation can improve performance by making the structure of harmful intent more explicit. Overall, ThreatCore provides a more consistent benchmark for studying fine-grained threat detection and highlights the challenges that current models still face in identifying indirect expressions of harmful intent.
Download

Paper Nr: 178
Title:

Multi-Group Fairness Measures for Classification

Authors:

Massimiliano Mancini, Donatella Merlini and Maria Cecilia Verri

Abstract: This paper proposes a robust, reliable, and reproducible methodology for evaluating group fairness in classification algorithms. Building upon established theoretical definitions of non-discrimination criteria Independence, Separation, and Sufficiency- we present a comprehensive approach to quantifying fairness. Notably, our methodology extends beyond binary group comparisons to accommodate scenarios with multiple sensitive groups. Furthermore, we provide a methodological approach for assessing fairness within the Sufficiency criterion by operationalizing Calibration, elucidating critical issues and conceptual subtleties that appear to have been overlooked in existing literature. We assess the reliability and robustness of our model through applications to one real dataset.
Download

Paper Nr: 182
Title:

When Fake-News Fakes: Foiling Feature-Based Attacks on Algorithmic Fake-News Detection

Authors:

Rany Zeidan, Gal Neria and Joachim Meyer

Abstract: The rapid spread of fake news threatens public trust and societal stability. Machine learning (ML) models have demonstrated strong performance in detecting fake news; however, they remain vulnerable to adversarial manipulations, specifically subtle changes that deceive classifiers. We present these vulnerabilities using the WELFake, ISOT, and Jruvika fake news datasets. We introduce a targeted ”feature hacking” approach that systematically manipulates key features such as sentiment polarity, punctuation, and capitalization to expose weaknesses in detection models. Using explainability tools such as SHapley Additive exPlanations (SHAP), we identify the most influential features and reveal how adversarial attacks exploit them. To address these manipulations, we evaluate adversarial retraining, in which models are retrained on manipulated examples to enhance their resilience to attacks. We show that relatively limited retraining with the altered data can effectively counteract manipulations. Our evaluation across the three datasets demonstrates significant improvements in both accuracy and robustness under diverse adversarial scenarios. This work underscores the need for fake news detection systems that are not only accurate but also resilient and adaptable. By systematically exposing vulnerabilities in established features and empirically evaluating practical defenses, we provide a pathway toward more robust and interpretable solutions against misinformation.
Download

Paper Nr: 133
Title:

Development of a Blockchain-Based Online Review System to Enhance Corporate Reputation

Authors:

R. E. Loke and J. Schoonderbeek

Abstract: The trustworthiness of online reviews is often questioned in e-commerce and this has a negative effect on the corporate reputation of organizations. The capabilities of blockchain technology could solve some reliability issues, for example regarding transparency and trustworthiness, and take away the need for a central database (Swan, 2015). With the aim to enhance corporate reputation for organizations, our study addresses these issues by applying an integrated blockchain-based online review system that stores online reviews of products on IPFS and computes, updates and retrieves actual, confident reputation scores and IPFS CIDs for products on the Ethereum blockchain. Zhou et al. (2021) created a similar system, nevertheless, without explicitly using sentiment analysis when computing the reputation score. Results on an Amazon showcase of a healthcare product shows that our novel reputation score is more reliable and resistant against misinterpretations.
Download

Paper Nr: 174
Title:

Privacy-Preserving User Authentication Architecture for Secure Human Interaction in Industrial AIoT Systems

Authors:

Ourania Manta, Angelos Mavrias and Andreas Miaoudakis

Abstract: Secure and accountable operation of industrial Artificial Intelligence of Things (AIoT) systems increasingly depends on robust identity and access management mechanisms that enable controlled human interaction with distributed runtime components. Despite growing emphasis on trustworthy AI, authentication and authorisation infrastructures are often treated as peripheral implementation details rather than foundational architectural elements. This paper presents a unified, Keycloak-based authentication architecture designed to support secure, role-aware and privacy-preserving human interaction within industrial AIoT pipelines. The proposed approach introduces a horizontally propagated token model in which JSON Web Tokens containing minimal identity, role and organisational claims are issued upon authentication and independently validated across user interface, orchestration and backend services. This stateless propagation strategy enables consistent access control enforcement at integration boundaries without requiring centralised session management. Multi-tenant isolation and privacy-by-design principles are embedded directly into identity claims and configuration loading mechanisms, aligning with governance requirements such as the General Data Protection Regulation. Deployment in representative industrial environments demonstrates the feasibility of the approach as an incrementally deployable authentication integration approach for regulated, multi-tenant AIoT ecosystems. The contribution lies in an architectural security integration model that bridges identity federation, role-based access control and AIoT runtime orchestration under real-world operational constraints.
Download

Area 5 - Data Science & Machine Learning

Full Papers
Paper Nr: 44
Title:

Customer Churn Prediction in the Telecommunications Sector Using Explainable Artificial Intelligence

Authors:

David Veríssimo, Joana Leite and Maryam Abbasi

Abstract: Customer churn poses a significant threat to profitability in the saturated telecommunications industry, yet accurately predicting churn remains challenging due to high-dimensional data and class imbalance where churners represent a small minority. While complex ensemble methods achieve high accuracy, their ”black-box” nature limits business adoption, as practitioners require transparent insights to design effective retention campaigns. This paper proposes a comprehensive Machine Learning pipeline that bridges the gap between predictive performance and model interpretability. We implement and compare five classifiers-Logistic Regression, Random Forest, XGBoost, LightGBM, and CatBoost-optimized via Bayesian hyperparameter tuning and evaluated using the recall-focused F2-score to address class imbalance. Our results demonstrate that gradient boosting models, particularly XGBoost, outperform aggregation-based ensemble strategies, achieving the highest F2-score of 0.7500 with a recall of 0.9385. Crucially, we integrate SHAP-based Explainable Artificial Intelligence to provide both global and local interpretability, revealing that contract type, tenure, and technical support subscriptions are the primary churn drivers. This dual-layer transparency enables targeted retention strategies while preserving high recall in imbalanced telecom data.
Download

Paper Nr: 48
Title:

Evaluating Heterogeneous Graph Neural Networks for Fashion Retail Recommendation: A Comparative Study on the H&M Dataset

Authors:

Davi Rocha de Santana, Leila Weitzel, Anderson Amendoeira Namen and Graziela Ferreira Guarda

Abstract: Recommender systems operating on large-scale e-commerce data must capture complex, multi-relational user–item interactions that conventional methods frequently fail to represent adequately. This study presents a comparative empirical evaluation of three recommendation architectures applied to the H&M retail dataset (~6 million transactions, RelBench): A Heterogeneous Graph Neural Network (HGNN) with dual Graph Attention layers, a Two-Tower neural model (2TT), and Neural Collaborative Filtering (CF). All models are evaluated under a rigorous user-level ranking protocol (1 positive + 99 negative candidates), using Mean Average Precision at 10 (MAP@10), Normalized Discounted Cumulative Gain at 10 (NDCG@10), and Area Un-der the ROC Curve (AUC) as primary metrics. HGNN achieves the highest ranking performance (MAP@10 = 0.1215; NDCG@10 = 0.1667; AUC = 0.7842), yielding relative improvements of approximately 29% over CF and 13% over 2TT in MAP@10. These results provide empirical evidence that heterogeneous graph-based representations, by encoding higher-order relational structures through message-passing over typed edges, yield measurable gains in top-k ranking precision compared to pairwise embedding methods. The findings also highlight an important distinction: global separability metrics such as AUC may not reliably predict top-ranked recommendation quality, underscoring the relevance of MAP and NDCG as primary decision criteria in this setting.
Download

Paper Nr: 58
Title:

Structural-Temporal Anomaly Detection in Scholarly Networks Using Hyperbolic Representations

Authors:

Renata Avros, Shoval Ben Shushan, Adar Budomski, Dvora Toledano-Kitai and Zeev Volkovich

Abstract: Dynamic citation networks reflect the hierarchical organization and temporal evolution of scientific influence. Detecting anomalous citation behavior within such networks remains challenging due to their scale-free structure, inherent hierarchy, and continuous growth. This paper introduces a Lorentzian Anomaly Attention framework, a structural–temporal model that integrates hyperbolic embedding in Lorentz space with masked temporal self-attention. The proposed approach jointly captures hierarchical geometry and evolving citation dynamics. Node representations are evaluated using unsupervised anomaly detection techniques. To compensate for the absence of ground-truth anomaly labels, a controlled synthetic injection protocol with statistically measurable evaluation criteria is developed. Experimental results demonstrate strong robustness under structural perturbations and indicate that global anomaly detection methods outperform local density-based approaches in hyperbolic dynamic environments. A theoretical analysis further provides geometric and spectral insight into anomaly separability within Lorentz space.
Download

Paper Nr: 59
Title:

Deep Representative Text Modeling for the Consistent Identification of SARS-CoV-2 Anomalies

Authors:

Renata Avros, Valery Kirzner, Dvora Toledano-Kitai and Zeev Volkovich

Abstract: The COVID-19 pandemic has emphasized the critical need for rigorous quantitative approaches to analyzing complex evolutionary systems. The identification of anomalous evolutionary trajectories, particularly within viral families such as SARS coronaviruses, offers essential insights into diversification mechanisms and facilitates the prediction of emergent variants. This work introduces a generalized Deep Impostor framework for detecting structural anomalies in large-scale viral datasets, grounded in the paradigm of Deep Text Modeling. By treating genomic sequences as complex texts written in the language of nucleotides, the methodology employs deep learning classifier ensembles trained on ”impostor” sample pairs to identify deviations from expected behavior. Viral sequences are systematically segmented and transformed into functional text representations using Convolutional Neural Networks, followed by clustering based on derived anomaly scores. Furthermore, the formal statistical consistency of this text-based anomaly mechanism is established under weak dependence assumptions. Experimental analyses demonstrate the framework’s robustness in uncovering latent evolutionary dependencies, specifically highlighting potential cross-border transmission vectors between geographically disparate regions, including the USA, Iraq, and Brazil.
Download

Paper Nr: 61
Title:

Community-Aware Spectral Anomaly Detection in Scholarly Graphs

Authors:

Itay Mohabati, Shai Yosef, Renata Avros and Zeev Volkovich

Abstract: Citation-based metrics are increasingly influencing academic evaluation, incentivizing manipulative practices that undermine bibliometric integrity. This paper presents a framework for detecting anomalous citations in scholarly networks by integrating graph neural network embeddings with community-aware spectral filtering. The framework uses GraphSAGE and Graph Convolutional Networks approaches to generate high-dimensional node representations. These embeddings are integrated with clustering techniques to delineate the network’s latent community structure. Building on this modular foundation, the SpecF algorithm leverages the Graph Fourier Transform and a community-aware low-pass spectral filter to isolate structural anomalies and detect nodes that deviate from community-specific smoothness patterns. Theoretical analysis establishes that anomaly detectability depends on high-frequency spectral energy rather than node signal amplitude, with optimal performance at moderate perturbation intensities. Experimental evaluation on the Cora and PubMed citation networks demonstrates effective anomaly identification under controlled conditions. The framework provides a principled approach to context-aware anomaly detection in citation networks, enabling the identification of locally irregular patterns overlooked by global statistical methods.
Download

Paper Nr: 64
Title:

Imputation Strategies for Predicting Tropospheric Ozone

Authors:

Miraceli Bonjardim Waldemar, Bruno Nazario Alves, Fernando Augusto Silveira Armani and Yuri Kaszubowski Lopes

Abstract: Tropospheric Ozone (TO) prediction is essential for air-quality management; however, the construction of robust and accurate predictive models is hindered by missing data in Air-Quality Station (AQS) time series. The development of reliable predictive models depends on uninterrupted datasets to properly capture the underlying dynamics and ensure stable parameter estimation. AQS time series frequently contain missing data due to sensor malfunctions, maintenance activities, power interruptions, communication failures, and equipment disruption caused by wildlife. This study evaluates two imputation strategies, Multivariate Imputation by Chained Equations (MICE) and the incorporation of data from nearby Weather Stations (WS), to mitigate missing data in AQS time series. Using multi-year real data from an AQS in Paraná, Brazil, we compare the performance of XGBoost models trained with each imputation strategy for TO prediction. The results show that, although both methods are robust, imputation using nearby WS data yields similar performance (R² = 0.87, RMSE = 2.60 ppb) compared to MICE (R² = 0.86, RMSE = 2.66 ppb). These findings suggest that different imputation strategies can yield comparable performance when developing predictive models for TO.
Download

Paper Nr: 66
Title:

An Optimized Ensemble Framework with Explainable AI for Proactive Credit Card Fraud Detection in Banking

Authors:

Henrique Barros, Francisco Antunes and Maryam Abbasi

Abstract: This study proposes an optimized ensemble framework for highly imbalanced credit card fraud detection that jointly maximizes predictive sensitivity and model interpretability. While machine learning offers significant potential for high-dimensional financial data analysis, its adoption is often hindered by extreme data sparsity and the inherent “black-box” nature of complex ensemble architectures. To bridge this gap, we develop a comprehensive pipeline incorporating six high-performance classifiers, fine-tuned via the Optuna Bayesian optimization framework. Leveraging a benchmark European dataset with a 0.172% fraud rate, our methodology employs a hybrid strategy of synthetic oversampling and cost-sensitive learning, integrated with SHapley Additive exPlanations (SHAP) to deconstruct complex decision boundaries. Empirical findings indicate that XGBoost delivers a superior Precision-Recall Area Under the Curve (PR-AUC) of 0.8755, while the Random Forest (RF) model achieves an optimal F2-Score (0.8673), effectively minimizing undetected fraudulent events. Our analysis further identifies time-of-day signals in the engineered temporal representation as relevant indicators of malicious activity. This work provides a mathematically grounded benchmarking framework for integrating Explainable Artificial Intelligence (XAI) into fraud detection pipelines, aligning high-accuracy analytics with the transparency requirements expected in regulated financial environments.
Download

Paper Nr: 86
Title:

Continuous Operation of a Myoelectric Hand Prosthesis Using Reinforcement Learning

Authors:

Carlos Ayala-Tipan, Jose-Fuertes Bustos, Lorena Isabel Barona López, Ángel Leonardo Valdivieso Caraguay and Marco E. Benalcázar

Abstract: Achieving continuous operation of prosthetic hands from electromyographic (EMG) features remains a significant challenge in assistive robotics, where traditional pattern recognition approaches struggle with real-time continuous operation requirements. This work presents a Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning algorithm trained to map EMG features to motor commands for a 3D-printed prosthetic hand. The agent architecture and reward structure underwent iterative refinement through experimental observations to optimize continuous operation performance. The agent was trained on EMG data collected from 12 participants at an academic research laboratory and evaluated in simulation and on physical hardware. In simulation, it achieved a mean absolute error between 0.07 and 0.28 when tracking flexion-glove–derived reference trajectories, while hardware testing revealed finger actuation capabilities. These results establish a foundation for exploring deep reinforcement learning approaches in EMG-based prosthetic operation.
Download

Paper Nr: 88
Title:

SecureBreak: A Dataset towards Safe and Secure Models

Authors:

Marco Arazzi, Vignesh Kumar Kembu and Antonino Nocera

Abstract: Large language models are becoming pervasive core components in many real-world applications. As a consequence, security alignment represents a critical requirement for their safe deployment. Although previous related works focused primarily on model architectures and alignment methodologies, these approaches alone cannot ensure the complete elimination of harmful generations. This concern is reinforced by the growing body of scientific literature showing that attacks, such as jailbreaking and prompt injection, can bypass existing security alignment mechanisms. As a consequence, additional security strategies are needed both to provide qualitative feedback on the robustness of the obtained security alignment at the training stage, and to create an “ultimate” defense layer to block unsafe outputs possibly produced by deployed models. To provide a contribution in this scenario, this paper introduces SecureBreak, a safety-oriented dataset designed to support the development of AI-driven solutions for detecting harmful LLM outputs caused by residual weaknesses in security alignment. The strong reliability of the proposed dataset derives from the adopted manual annotation procedure, in which labels are assigned conservatively to prioritize safety even in the presence of minor disagreements in the annotators’ opinion. Our exploratory data analysis campaign shows satisfactory performance in the detection of unsafe content across several risk categories. To evaluate its effectiveness, we measure the performance of several pre-trained LLMs in the considered classification setting under baseline conditions and compare these results with those obtained after fine-tuning the same models on SecureBreak. The results indicate that the dataset is valuable not only for constructing post-generation filtering modules that act as a last-line defense, but also for building additional supervisory intelligence for alignment optimization. We show that models fine-tuned on SecureBreak improve safety classification by up to 20% over baselines, with small models reaching up to 90.14% accuracy and sometimes outperforming larger models. In particular, classifiers derived from SecureBreak can be used to measure residual safety failures, inform whether additional training or refinement steps are necessary, and ultimately support more controlled and effective security alignment workflows.
Download

Paper Nr: 97
Title:

Prompting vs Ensemble Architectures for Arabic English Code-Switched Classification

Authors:

Donia Ali, Salma Haytham, Sandra George and Caroline Sabty

Abstract: Arabic English code-switching is common in online communication, where users mix both languages within the same sentence. Mixed scripts, dialectal Arabic, Arabizi, and inconsistent spelling create instability for NLP models and make tasks such as topic classification difficult. Previous work has typically focused either on improving transformer architectures through ensembles and multi view modeling, or on enhancing inference through prompt design in large language models. However, these directions are rarely evaluated under the same experimental setup. In this paper, we provide a controlled comparison between architecture based optimization and prompting based inference for Arabic English code-switched topic classification. Architectural experiments include single model baselines, voting ensembles, stacked meta learning, and translation based multi view inference. Prompting experiments evaluate zero shot, few shot, retrieval augmented generation (RAG), and reasoning based strategies using LLaMA 3.3 70B without fine tuning. While architectural extensions progressively improve performance, reaching 0.92 Macro F1 with translation based multi view and Random Forest, retrieval augmented few shot prompting achieves 0.98 Macro F1, surpassing all ensemble configurations. The results indicate that inference time conditioning can be more effective than increasing architectural complexity for this task.
Download

Paper Nr: 99
Title:

ResSET: Enhanced Motor Imagery Classification Using Hybrid Transformer-ResNet Architectures and Generative Data Augmentation

Authors:

Duc Thien Pham and Roman Mouček

Abstract: Brain-computer interfaces (BCIs) have the potential to bring about revolutionary changes in various fields, e.g., healthcare, rehabilitation, and human augmentation. In the healthcare sector, BCIs are making remarkable advancements, particularly in aiding rehabilitation for individuals with motor impairments through analyzing motor imagery (MI) using electroencephalograms (EEG). Traditional approaches often struggle to achieve high accuracy due to the complexity and variability of MI signals, requiring more complex deep-learning models to improve classification performance. This study introduces ResSET, a hybrid Residual Squeeze-Excitation Transformer network designed to optimize MI classification for both binary and multiclass tasks. To overcome signal variability and data scarcity, we employed three augmentation strategies: Noise Injection (NI), conditional variational autoencoder (cVAE), and Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (cWGAN-GP). The framework’s efficacy was validated across three public datasets: MI-EEG (29 subjects), BCI Competition IV-2a (9 subjects), and the EEGMMIDB dataset (103 subjects). The results demonstrate the effectiveness of the proposed model, achieving competitive performance compared to previously reported state-of-the-art studies in classifying MI using multi-channel EEG signals.
Download

Paper Nr: 106
Title:

ContrastSkill: Task-Oriented Contrastive Intermediate Training for Skill Extraction

Authors:

Aleksander Bielinski, David Brazier, Dimitra Gkatzia and Alistair Lawson

Abstract: Rapidly evolving global labor markets require scalable workforce intelligence systems for precise skill extraction from unstructured text. While modern transformer-based architectures provide a foundation for this task, their efficacy is constrained by the data-annotation bottleneck. Acquiring expert labeled data across a shifting skill landscape is prohibitively expensive. We introduce ContrastSkill, a multi-stage framework that leverages supervised contrastive intermediate training to enhance skill extraction robustness and generalization. Unlike traditional transfer learning, our approach utilizes abundant, low-cost weak labels derived from rule-based matching to shape a task-aligned representation space before final fine-tuning on scarce expert annotations. We evaluate ContrastSkill across three diverse benchmarks: GREEN, SKILLSPAN, and SAYFULLINA, and compare its performance against traditional encoder models, domain-adapted variants like JobBERTa, and Large Language Models (LLMs) such as GPT-4.1. Our results show statistically significant improvements across multiple benchmarks, including a 2.32 percentage point Span-F1 improvement in the cross-dataset setting. Benchmarking further reveals that LLMs exhibit high recall but struggle with precision in exact span delineation. The framework requires no additional expert annotation beyond existing benchmark datasets, relying solely on inexpensive rule-derived signals. We release our code, matching rules, and trained models to support reproducibility.
Download

Paper Nr: 155
Title:

Enabling Efficient Domain Adaptation via Noise-Enhanced Flow Matching

Authors:

Aitian Ma, Dongsheng Luo and Mo Sha

Abstract: Domain adaptation remains a significant challenge in deploying data-driven models under distribution shifts, particularly when transferring from simulated to real-world environments. Existing approaches often rely on large labeled target datasets, suffer negative transfer, and provide limited interpretability. In this paper, we present NoiseFlow, a data-efficient domain adaptation framework that leverages noise-aware modeling and flow matching to enable robust cross-domain generalization. Our key insight is that feature dimensions exhibit heterogeneous sensitivity to noise, which can be amplified under domain shift. NoiseFlow introduces a feature-aware teacher student architecture that combines knowledge distillation, distribution alignment, and continuous flow matching to learn smooth transformations between source and target domains. Experimentation on wireless network configuration tasks demonstrates that NoiseFlow achieves good performance in low-data regimes, reaching 69.8% accuracy with a single target sample and improving zero-shot transfer performance by up to 40% over existing methods.
Download

Paper Nr: 161
Title:

A Faithful Multi-Head Framework for Explainable Toxicity Detection with Structured LLM-Based Synthetic Supervision

Authors:

Linh Nguyen and Tran Anh Tuan

Abstract: Detecting toxic speech on social media requires not only accurate classification but also interpretable explanations to support moderation decisions. This challenge is particularly important for Vietnamese, where fine-grained annotated resources remain limited. In this work, we propose a Multi-Head Explainable Framework that jointly predicts toxicity labels together with structured rationales, including Act, Target, and Evidence Spans. To address annotation scarcity, we employ a structured synthetic supervision strategy based on Large Language Models, combined with a self-check mechanism and Hard Sample Mining to improve annotation consistency and data quality. The model is trained through a three-phase curriculum designed to transfer knowledge from synthetic rationale supervision to real-world Vietnamese social media data. Experimental results on the ViHSD benchmark show that the proposed framework improves toxicity detection performance and reduces false positives compared with baseline models, achieving a peak F1toxic score of 0.7858 with the Qwen3-8B backbone. In addition, the framework provides faithful explanations, with ViHateT5-base achieving the highest Faithfulness score of 0.7824. Qualitative analysis further shows that the model can identify explicit evidence spans associated with toxic content, although challenges remain in handling sarcasm and implicit toxicity. These findings suggest that structured rationale supervision is a promising direction for improving both detection performance and interpretability in Vietnamese toxicity detection.
Download

Paper Nr: 165
Title:

Evaluating Local Explainability Metrics for Machine Learning Models on Tabular Data

Authors:

Tomás Pereira, João Vitorino, Eva Maia and Isabel Praça

Abstract: Despite the wide use of explainability techniques to attempt to understand the behavior of Artificial Intelligence (AI), the generated explanations may not always be reliable. An explanation can appear plausible to humans but fail to capture the internal reasoning of a model, particularly when dealing with complex tabular data. This paper studies the trustworthiness of local explainability techniques when applied to complex tabular classification tasks, considering evaluated metrics for three main properties: faithfulness to the model’s predictions, robustness to input data variations, and complexity of the explanation itself. A benchmark was performed for Local Interpretable Model-Agnostic Explanations (LIME), Kernel SHapley Additive exPlanations (SHAP), and Feature Ablation techniques, across 32 datasets and different types of machine learning models. Model performance ranges were analyzed to identify two groups: consensus-correct, which are samples that all models predicted correctly, and consensus-wrong, samples that all models predicted incorrectly. The obtained results demonstrate that that the explanations are not always correlated with a model’s predictive performance. Instead, dataset complexity and feature distributions seem to be the main factors affecting explanation quality and reliability.
Download

Paper Nr: 169
Title:

Actionable Explainable AI for Credit Risk: A SAFE-AI Framework Integrating SHAP and Counterfactual Explanations with DiCE

Authors:

Matheus Francelino Bezerra da Silva and Carlos Henrique Quartucci Forster

Abstract: The increasing use of machine learning models in credit risk assessment raises concerns about transparency, trust, and regulatory compliance, as many high-performing models behave as black boxes. This paper proposes an actionable explainable AI framework for credit risk that combines an eXtreme Gradient Boosting (XGBoost) classifier with Shapley Additive Explanations (SHAP) and Diverse Counterfactual Explanations (DiCE), organized under the Situation Awareness Framework for Explainable AI (SAFE-AI). Using the public HELOC dataset, we first train and tune an XGBoost model to predict default risk, then derive global and local explanations with SHAP, and finally generate counterfactual explanations with DiCE to indicate feasible changes capable of reversing unfavorable outcomes. Results show that the proposed framework provides competitive predictive performance while enhancing interpretability and actionable decision support in credit risk assessment.
Download

Paper Nr: 176
Title:

Dysgraphia Detection from Handwritten Text Using a Hybrid ConvNeXt-Transformer Model

Authors:

Wallace Primo Freire Junior, Simone Bello Kaminski Aires and Maria Eduarda Guedes Pinto Gianisella

Abstract: Dysgraphia is a learning disorder that impairs handwriting production, affecting the academic performance and development of school-aged children. Its diagnosis is typically conducted by specialists through qualitative assessment, a time-consuming process subject to variability among evaluators. This work proposes a hybrid architecture for the automatic identification of dysgraphic patterns in children’s handwriting, combining the ConvNeXt-Tiny convolutional backbone for local feature extraction with a Transformer Encoder for modeling global spatial dependencies, followed by a binary classification head based on CLS tokens. The pipeline incorporates a custom preprocessing filter, called WhiteInkAttentionThreshold, designed to isolate handwritten traits and suppress background noise present in the digitized images. The experiments were conducted on a dataset of 249 images from 83 participants, independently divided by participant. The proposed model achieved an accuracy of 94.59%, a macro F1-score of 94.49%, a macro precision of 95.45%, and a macro recall of 94.12% in the test set, surpassing the evaluated reference models. The ablation study demonstrated that the integration of the customized filter provided a gain of 6.05 percentage points in accuracy. The interpretability analysis via Grad-CAM confirmed that the model bases its decisions on structurally relevant regions of the handwritten trace, offering qualitative support for the validity of the reported metrics.
Download

Paper Nr: 181
Title:

Optimising Wheel Replacement Schedules for Freight Trains Based on Machine Learning Models

Authors:

Max Strang, Shaibal Barua, Mobyen Uddin Ahmed and Shahina Begum

Abstract: Efficient maintenance planning is crucial for railway operators to minimise costs and operational disruptions. Railway freight providers currently replace train wheels either during scheduled maintenance or as reactive repairs, leading to inefficiencies. This paper explores the application of machine learning to predictive maintenance by analysing sensor data from Dynamic Pressure Check (DPC) detectors mounted along railway tracks. These detectors measure the force impacts as the train moves, providing key indicators of wheel wear. Here, machine-learning-based predictive models with uncertainty quantification are developed to detect damaged train wheels. The model achieved R^2>0.98 with RMSEs of 0.08 kilonewton for the left wheel and 0.36 kilonewton for the right wheel. Further, forecasting trends over a three-year horizon identified one axle requiring replacement, with left wheels typically failing first. A maintenance scheduling strategy was implemented in which both wheels on an axle are replaced when either wheel exceeds predefined thresholds. This work demonstrates the potential of predictive maintenance to improve efficiency and reduce costs in railway operations by applying eXtreme Gradient Boosting (XGBoost), Isolation Forest, and Linear regression methods to historical data from DPC detectors.
Download

Paper Nr: 184
Title:

A Two-Level Machine Learning Based-Intrusion Detection System for IoT Healthcare Application Based on Blockchain

Authors:

Manel Boujelben and Manel Hentati

Abstract: Smart healthcare systems offer human-centric solutions that enable the remote monitoring of patients, particularly those who are elderly, disabled, or located in geographically remote regions, thereby enhancing the quality and accessibility of medical services. These systems leverage core technologies such as the Internet of Medical Things (IoMT), blockchain, and artificial intelligence to facilitate the analysis and secure sharing of medical data among various stakeholders in the healthcare ecosystem. However, the transmission of sensitive health information over public networks raises significant security and privacy concerns. To address these issues, we propose a role-based access control protocol that restricts unauthorized access to the Ethereum blockchain and enforces rules for data usage. In addition, cryptographic primitives are employed to ensure data confidentiality. Our security framework also incorporates a two-level machine learning-based Intrusion Detection System (IDS): the first operates at the IoT gateway level to monitor IoT devices traffic, while the second is integrated within the blockchain network to detect and prevent malicious Ethereum transactions. Experimental evaluation on the Edge-IIoT and Ethereum fraud datasets demonstrates that the proposed IDS achieves high effectiveness across key metrics as accuracy, precision, recall, and F1-score. Random Forest outperforms all other algorithms, with accuracy rates of 97.12% for inside IDS and 98.2% for outside IDS. A comparison with state-of-the-art solutions demonstrates that our approach outperforms existing methods in both detection accuracy and security. Security analysis further confirms the system’s robustness against diverse cyberattacks.
Download

Paper Nr: 189
Title:

Query Adaptive Rank Fusion: A Training Free per Query Weighting Scheme for Sparse–Dense Hybrid Retrieval

Authors:

Mohammad Saad Rafeeq, Chandramani Chaudhary, Nirmal Kumar Boran and Ashok Kumar Suhag

Abstract: Reciprocal Rank Fusion (RRF) is the standard recipe for combining a sparse and a dense retriever in retrieval augmented generation, but it assigns both components a fixed and equal weight on every query. This is a strong assumption, since a query of rare specific tokens is best served by the sparse retriever while a paraphrastic query (one composed largely of common words that admits many surface forms of the same intent) is best served by the dense one. We show on four BEIR corpora that equal weight RRF fails to surpass its stronger component on any of them, and we propose Query Adaptive Rank Fusion (QARF), a training free variant whose fusion weight is a closed form function of standard corpus IDF statistics already available at index build time. QARF operates in two regimes selected automatically by the data: per query weighting on corpora with a broad IDF distribution, and automatic corpus level recalibration of RRF on corpora where almost every query is more specific than the mean. Using a BM25 + E5 large backbone, QARF improves over its own RRF baseline on all four corpora by +1.74, +5.12, +8.61, and +2.97 NDCG@10 points on NFCorpus, SciFact, FiQA, and ArguAna, matches pure E5 within one point on SciFact and FiQA, and exceeds it on ArguAna while retaining the hybrid’s wider candidate pool. End to end evaluation with Gemma 3 1B shows that the retrieval gain transfers downstream as a lower abstain rate on FiQA (4.6%→3.2%, the lowest of any retriever); on the other corpora the rate stays within noise of RRF, so we report commit rate as a per corpus companion metric rather than a uniform claim.
Download

Paper Nr: 198
Title:

Explaining Rankings with Prototypes and Criticisms

Authors:

Lars Schütz and Korinna Bade

Abstract: Explaining rankings is increasingly important in information retrieval, recommender systems, and other decision-support applications. Existing approaches often rely on feature attribution, surrogate models, or counterfactual explanations, communicating through feature weights, local approximations, or hypothetical changes rather than concrete examples. We propose a novel example-based approach that explains rankings through prototypes and criticisms. Prototypes are representative ranked items that summarize dominant relevance patterns, whereas criticisms highlight items insufficiently captured by those prototypes. We introduce the method rank-clust-cen, which selects prototypes and criticisms from a clustering of the top-k ranked items while preserving the original ranking order. In a controlled user study on real-world digital public participation data, participants inferred which candidate query generated a ranking under different explanation strategies. Our method achieved accuracy comparable to showing the full ranking without explanations while reducing task completion time and perceived difficulty, indicating that compact example-based explanations can improve the usability of ranking systems without sacrificing decision quality.
Download

Paper Nr: 207
Title:

Deep Learning on the Wing: A Survey of Resource-Efficient Object Detection and SLAM for Autonomous UAVs

Authors:

Rayan Abri, Sara Abri and Salih Cetin

Abstract: The evolution of Unmanned Aerial Vehicles (UAVs) into self-governing robotic entities is currently limited by the high computational overhead of modern neural networks. This review investigates the intersection of high-fidelity deep learning and the stringent resource limitations of edge-based aerial hardware. We present a systematic classification of recent breakthroughs in real-time perception, specifically evaluating stereo image processing and Simultaneous Localization and Mapping (SLAM) through the lens of computational economy. Drawing on extensive industrial experience in drone manufacturing and AI department leadership, this paper analyzes the efficacy of specialized optimization frameworks-such as multi-threaded frame tiling and hardware-concurrency mapping-designed to maximize inference speed on embedded CPU/GPU architectures. We further examine the role of spatio-temporal modeling and LSTM-based architectures in navigating unpredictable environments, while synthesizing the requirements for safety-critical, responsible AI deployment. By aligning theoretical algorithmic pruning with the practical realities of the product lifecycle, this survey provides a definitive technical roadmap for engineers and researchers aiming to achieve robust, on-board autonomy in the next generation of intelligent flight systems.
Download

Paper Nr: 222
Title:

SIGMA: A Lightweight Multi-Perspective Statistical-Signature Framework for Concept Drift Detection in Business Process Event Logs

Authors:

Lerina Aversano, Felice Franchini and Debora Montano

Abstract: Concept drift undermines the reliability of process mining: once the process generating an event log shifts, discovered models grow stale and conformance results become misleading. Current detectors lean either on deep neural models, data-hungry and opaque, or on single-perspective statistics blind to subtle cross-view changes. This paper introduces SIGMA, a lightweight training-free framework comparing four complementary statistical signatures (activity histograms, directly-follows graph edges, inter-event durations, case lengths) across consecutive sliding windows through non-parametric two-sample tests. A 2-of-4 voting rule fires once multiple perspectives concurrently signal a shift. The study evaluates seven configurations on (i) a synthetic generator with controlled sudden, gradual, incremental, and recurring drifts, (ii) the public Italian Road Traffic Fine Management (RTFM) log (561 470 events, 11 activities, 2000 to 2013), and (iii) Business Process Intelligence Challenge (BPIC) 2015 Municipality 1 (52 217 events, 398 activities, 2010 to 2015). On the synthetic benchmark, SIG-KS/χ2 attains 40% lower false-positive rate (FPR) than ProDrift. On the two real logs the picture grows more nuanced: ProDrift and naive baselines collapse into permanent-alarm detectors on RTFM (98.9% alarm rate), whereas the neural baseline over-fires on BPIC 2015 (70%). SIGMA holds comparatively stable alarm behaviour across both settings (58.1% on RTFM, 8.0% on BPIC 2015). The framework runs in milliseconds on a single CPU core.
Download

Paper Nr: 223
Title:

Probabilistic Variational Archetypal Learning on the Simplex for Interpretable Medical Image Analysis

Authors:

Ali Yassin, Narjes Soudin, Abbass Rammal and Mageda A. A. Sharafeddin

Abstract: Interpretable image learning has advanced through prototype- and concept-based neural models, but many methods explain predictions through local regions or predefined concepts rather than through a global geometric representation. Archetypal analysis provides such a representation by expressing each image as a convex mixture of a small set of extremal archetypes. However, deep archetypal learning remains underexplored in medical imaging. Deep Archetypal Analysis (DeepAA) introduced a probabilistic variational framework with simplex-constrained latent representations and side information, originally demonstrated for facial-expression image reconstruction. In this work, we extend DeepAA from JAFFE facial-expression analysis to interpretable dermatoscopic lesion analysis on HAM10000, a seven-class skin lesion dataset. We adapt DeepAA for medical images using a convolutional RGB architecture, a supervised lesion-classification objective, a TensorFlow 2/Keras implementation, and a scalable stratified data pipeline. These modifications preserve DeepAA’s simplex-constrained geometry while enabling application to larger and more heterogeneous medical images. We evaluate reconstruction fidelity, simplex adherence, coefficient sparsity, semantic alignment, and interpolation behavior across K ∈ 3,5,7,9 archetypes. JAFFE serves as a controlled reference domain, while HAM10000 tests whether archetypal structure can capture clinically meaningful lesion extremes. Overall, this work positions DeepAA as a geometrically grounded framework for interpretable medical image representation learning beyond local prototype or concept-based explanations.
Download

Paper Nr: 231
Title:

Plywood Veneer Quality Classification Using CNNs with Squeeze-and-Excitation Blocks

Authors:

Cristian Gotardo, Andreia Marini, Marlon Marcon and André Roberto Ortoncelli

Abstract: Laminate quality assessment is essential in the wood industry, as it directly affects productivity and profitability. Although most CNN-based studies focus on detecting specific wood defects, to the best of our knowledge, no prior work has been found that classifies complete laminate images directly into predefined quality levels in a single step. This study proposes a novel CNN-based approach for classifying wood laminates into five quality levels (A–E) according to the PS 1-22 standard. By assigning quality directly from a single inference, the method avoids the traditional two-stage defect-detection-and-quality-estimation process, making it suitable for real-time industrial applications. Experiments were conducted using 4,400 laminate images, including 2,054 generated through data augmentation. Six CNN architectures were evaluated with and without an additional Squeeze-and-Excitation (SE) block. The best model, EfficientNet-B7 with an additional SE block, achieved an average accuracy of 0.826, outperforming all baselines. Moreover, the SE block improved results in 4 of the 6 CNNs, indicating that attention mechanisms are effective for this task. Most errors occurred between adjacent classes, supporting the practical viability of the proposed approach for industrial use.
Download

Short Papers
Paper Nr: 12
Title:

Evaluating England's Plastic Bag Reduction Strategies, 2015-2025

Authors:

Erdem Sarialioglu

Abstract: This paper evaluates the long-term effectiveness of England's single-use plastic carrier bag charge between 2015 and 2025. Using official annual data from the Department for Environment, Food and Rural Affairs (DEFRA) and the House of Commons Library, the study examines four indicators: total plastic bags distributed, bags per capita, gross proceeds from the charge, and donations to good causes. The analysis focuses on two policy moments: the introduction of the 5p charge for large retailers in October 2015, and the increase to 10p with extension to all retailers in May 2021. The findings show that the 2015 charge was followed by a major reduction in plastic bag use, with total bags distributed falling from approximately 8.5 billion in 2014 to around 2.2 billion in 2015-16 and continuing to decline in later years. Per-capita use also fell sharply, reaching single-digit levels after the 2021 policy update. The 2021 intervention appears to have reinforced an already established reduction rather than creating a second decline of similar magnitude. At the same time, lower bag use reduced gross proceeds and donations to good causes, creating a trade-off between environmental success and charitable revenue. Overall, the evidence suggests that England's charging system produced substantial and sustained behavioural change, although annual aggregate data limits causal precision.
Download

Paper Nr: 35
Title:

Selection Bias Correction in Retail Intelligence

Authors:

Spandan Ghose Chowdhury

Abstract: Retail intelligence often relies on monitoring popular, high-velocity products, potentially biasing economic indicators by ignoring the “long tail” of niche items. This simulation study investigates selection bias in inflation estimation and compares correction methods across diverse data-generating processes. Through 400 Monte Carlo replications spanning four scenarios—aligned step functions, smooth gradients, misaligned breaks, and polynomial relationships—we test the robustness of Inverse Probability Weighting (IPW) with five specifications against stratification with varying strata counts. Our findings reveal fundamental limits of weighting methods in retail long-tail contexts: stratification achieves superior performance in three of four scenarios, maintaining sub-0.04pp median error even when boundaries deliberately misalign with population breaks (116× advantage over IPW). However, IPW with spline propensity models wins under smooth polynomial relationships (median error 0.007pp vs. 0.013pp), demonstrating context-dependency. Critically, even an oracle IPW specification with perfect structural knowledge achieves 6.06pp error compared to stratification’s 0.008pp in step-function scenarios. This reflects violation of the Positivity Assumption—a fundamental causal inference requirement—rather than IPW methodological inferiority. When selection probabilities differ dramatically (90% vs. 1%), weighting methods operate outside their theoretical design envelope. These results demonstrate that stratification provides a safer engineering choice in retail long-tail distributions with severe positivity violations.
Download

Paper Nr: 36
Title:

Towards Fine-Grained Emotion Detection in Hebrew: Dataset Creation and Evaluation

Authors:

Natalia Vanetik, Shilat Haya Yosefi and Keturah Shlomo

Abstract: Fine-grained emotion detection for Hebrew remains underexplored due to the lack of publicly available annotated resources. We present HeMotion, a manually annotated and balanced Hebrew dataset for emotion classification based on Plutchik’s eight-emotion model, consisting of 670 sentences. We analyze the dataset’s lexical and structural properties and establish the first benchmark for this task using Hebrew transformer models, a dedicated Hebrew emotion model, and large language models with prompting. Experimental results show that transformer and LLM-based approaches outperform traditional baselines. By releasing HeMotion and the benchmark results, this work provides a new resource and evaluation framework for fine-grained emotion detection in Hebrew.
Download

Paper Nr: 39
Title:

Identification of Relevant Association Rules from Biased Target Value Distributions of Empirical Studies

Authors:

Friedemann Schwenkreis

Abstract: Empirical studies often face challenges with imbalanced data distributions, particularly when the value of interest of the target attribute occurs rarely. Traditional classification and association rule learning approaches generate an overwhelming number of rules, making manual identification of relevant rules impractical. This paper introduces the Double Sigmoid-based Relevance (DeSiRe) indicator, a novel measure combining mapped support and confidence values based on a logistic function to effectively identify rules indicating potential causal dependencies. Using a dataset of knee injuries among field hockey players, the DeSiRe indicator demonstrates improved identification of actionable rules, facilitating hypothesis generation for injury prevention. The approach enables researchers to prioritize rules for further evaluation, significantly reducing verification effort and enhancing insight extraction from empirical data.
Download

Paper Nr: 43
Title:

DeepSupp: Attention-Driven Correlation Pattern Analysis for Dynamic Time Series Support and Resistance Levels Identification

Authors:

Boris Kriuk, Logic Ng and Zarif Al Hossain

Abstract: Support and resistance (SR) levels are central to technical analysis, guiding traders in entry, exit, and risk management. Despite widespread use, traditional SR identification methods often fail to adapt to the complexities of modern, volatile markets. Recent research has introduced machine learning techniques to address the following challenges, yet most focus on price prediction rather than structural level identification. This paper presents DeepSupp, a new deep learning approach for detecting financial support levels using multi-head attention mechanisms to analyze spatial correlations and market microstructure relationships. DeepSupp integrates advanced feature engineering, constructing dynamic correlation matrices that capture evolving market relationships, and employs an attention-based autoencoder for robust representation learning. The final support levels are extracted through unsupervised clustering, leveraging DBSCAN to identify significant price thresholds. Comprehensive evaluations on S&P 500 tickers demonstrate that DeepSupp outperforms six baseline methods, achieving state-of-the-art performance across six financial metrics, including essential support accuracy and market regime sensitivity. With consistent results across diverse market conditions, DeepSupp addresses critical gaps in SR level detection, offering a scalable and reliable solution for modern financial analysis. Our approach highlights the potential of attention-based architectures to uncover nuanced market patterns and improve technical trading strategies.
Download

Paper Nr: 50
Title:

Enhancing Visual Perception in Foggy Conditions via Multiclass Fog Density Modeling

Authors:

Mohamad Mofeed Chaar and Galia Weidl

Abstract: Autonomous driving (AD) systems have advanced rapidly over the past decade; however, robust perception under adverse weather conditions remains a major challenge, particularly under dense fog conditions. In this work, we investigate fog-aware perception using synthetically generated fog data derived from the Waymo dataset. To support fog simulation, depth images are generated using an iterative learning approach. We consider five fog-density levels: clear weather, light fog, moderate fog, heavy fog, and very heavy fog. Instead of training a single unified model across all conditions, we train separate perception models for each fog-density level. Experimental results show that density-specific training improves performance under severe fog conditions. In particular, for the very heavy fog class, recall improves from 0.076 to 0.254, corresponding to an absolute gain of 17.8 percentage points. These findings suggest that deploying multiple specialized models-rather than a single general-purpose model-can improve perception robustness for autonomous vehicles under challenging visibility conditions. Future work will extend this strategy to additional sensing modalities, including LiDAR and radar, and evaluate generalization across diverse weather scenarios.
Download

Paper Nr: 52
Title:

Spatially Constrained Clustering for Analyzing Spatio-Temporal Dynamics of Automobile Insurance Risk

Authors:

Shengkun Xie, Samy El Mouttalibi, Jin Zhang and Clare Chua-Chow

Abstract: Territorial risk modeling in auto insurance presents a complex spatial clustering challenge, where economic, demographic, and driving behavioral shifts can rapidly change the distribution of insurance claims, including both claim frequency and severity. When analyzing claim loss data patterns, the use of conventional clustering methods disregards spatial dependencies, leading to spatially overlapped territories that are impractical for rate-making. This paper presents a computational framework using spatially constrained clustering that integrates geographic contiguity directly into the data partitioning process, producing rating territories that are both spatially coherent and statistically homogeneous. Using large-scale Automobile Statistical Plan (ASP) data from Ontario, Canada, the framework jointly analyzes three loss metrics, claim frequency, claim severity, and loss cost, across three major auto insurance coverages, including Accident Benefits (AB), Collision (CL), and Third-Party Liability (TPL). Experimental results demonstrate that the proposed method enables spatial coherence, interpretability, and adaptive territorial segmentation. The approach provides a scalable, data-driven solution to spatial risk modeling, with applications in actuarial data science, geospatial analytics, and other domains where interpretability and constraint-aware clustering are required.
Download

Paper Nr: 54
Title:

Hybrid Approach Based on Supervised Learning and Synthetic Data Generation for University Student Dropout Risk Prediction

Authors:

Laura Caicedo, Juan José Muñoz and Nestor Diaz

Abstract: Student dropout in higher education constitutes a structural problem that affects academic quality and institutional sustainability. In Colombia, between 30% and 50% of students left their studies, highlighting the need to strengthen early detection systems. However, the performance of supervised models is often affected by class imbalance, data scarcity, and constraints associated with the use of sensitive information. This study proposes a hybrid methodological approach structured under the CRISP-DM framework that integrates a supervised learning model with synthetic data generated by generative artificial intelligence. The workflow incorporates exploratory data analysis, business rules embedded in the generative process, statistical validation of synthetic data, and comparative evaluation under a no-data-leakage setting. The results show that incorporating synthetic data exclusively into the training set improves predictive performance, particularly for minority classes, reflected in gains in precision, recall, and F1-score, while preserving evaluation on untouched real test data. The proposed approach offers a reproducible, transferable methodological alternative to mitigate class imbalance in university student dropout prediction.
Download

Paper Nr: 55
Title:

Regime Detection and Forecasting of Financial Indicators in Electric Transmission Sector Companies Using Hidden Markov Models

Authors:

Giovanni Vallim Maniezzo, Elder Oroski and Luis Ledo Mota Melo Júnior

Abstract: The Brazilian electric transmission sector operates under a regulated revenue regime, yet remains subject to financial volatility arising from internal corporate strategies and external macroeconomic shocks. This study compares four Hidden Markov Model (HMM) variants-Categorical, Gaussian (GHMM), Gaussian Mixture (GMM-HMM), and Autoregressive (AR-HMM)-to estimate financial regimes of four transmission companies listed on the Brazilian stock exchange (B3), using quarterly data from 2010 to 2024. The regulated nature of this sector provides a controlled environment with reduced speculative noise, enabling the comparison of regime-detection accuracy across model variants. Among continuous-emission models, the GHMM achieved the lowest forecasting error (NRMSE between 0.170 and 0.221), while the Categorical HMM attained one-step-ahead accuracies up to 0.609, exceeding the random-guessing baseline. The AR-HMM failed to converge for three of the four companies within the 35-quarter training window. The decoded regimes suggest that revenue-based indicators are more strongly associated with firm-specific dynamics, whereas operational expenses exhibited the highest cross-company synchronization rate (13.793%), consistent with common inflationary pressures. Regime transitions were detected from Q4/2012 onward, a timing that is temporally consistent with the period following Provisional Measure No. 579/2012. These findings indicate that the GHMM is the most effective HMM variant within the data-sparse conditions examined.
Download

Paper Nr: 62
Title:

A Unified Semantic Representation Framework for Multi-Source Heterogeneous Data Fusing Protocol Syntax Constraints and Cross-Source Alignment: A Case Study of in-Vehicle Intrusion Detection

Authors:

Yixuan Liu, Cong Yang, Haibo Cheng, Kaimin Zhuo, Peng Liu and Yixuan Wang

Abstract: Unified semantic representation of multi-source heterogeneous data constitutes the fundamental basis for cross-source collaborative perception in complex cyber-physical systems. Existing methods broadly suffer from three interconnected deficiencies: protocol specification priors are neglected, alignment strategies remain confined to coarse-grained instance-level matching, and none systematically resolves the triple barriers of structural, temporal, and semantic gaps in strongly heterogeneous scenarios. This paper proposes a unified semantic representation framework integrating protocol syntax constraints with cross-source semantic alignment, comprising three tightly coordinated modules: a specification constraint modeling layer grounded in context-sensitive grammar, which quantifies structural, temporal, and semantic deviations through a three-dimensional orthogonal mechanism and injects constraint-aware vectors via cross-attention at each encoding step; a constraint-aware hierarchical Transformer encoding layer that decouples shared backbone and source-specific adapters with absolute physical timestamp embedding, outputting multi-granularity unified representations; and a cross-source semantic alignment layer that constructs Gaussian-weighted time-co-occurrence positive pairs and mines hard negatives from a momentum queue, optimized through multi-granularity joint contrastive loss and intra-source uniformity regularization. Experiments on 20,000 strongly heterogeneous samples demonstrate alignment of 0.541 and uniformity of -2.436, outperforming the strongest baseline by 35.1% and 0.783 respectively; mAP@1 reaches 74.6%, downstream anomaly detection F1 reaches 91.5% with AUC-ROC of 94.3%. Transfer validation on in-vehicle intrusion detection yields an F1-score of 95.2%, AUC-ROC of 97.1%, and a false alarm rate of only 2.2%. Ablation experiments confirm the independent and irreplaceable contribution of each module.
Download

Paper Nr: 74
Title:

Quantum Computed Moments Drive Orthogonal Neodymium Selectivity in Lanmodulin Dimer Pockets through Hamiltonian Distance Learning

Authors:

Don Roosan, Mazharul Karim, Rubayat Khan and Md Rahatul Ashakin

Abstract: Rare-earth separation remains difficult because adjacent lanthanides differ only subtly in size while behaving similarly in classical extraction systems. Neodymium is especially important for permanent magnet supply chains, yet current recovery routes remain chemically intensive. This study presents a computational framework that couples a reduced-order Quantum Computed Moments workflow with Hamiltonian-distance-guided design to engineer neodymium-selective lanmodulin dimer pockets. The workflow integrates the 8FNS lanmodulin-neodymium structure, coordination shell analysis, ligand-field and covalency descriptors, and an interpretable five-state reduced Hamiltonian spanning neodymium, promethium, samarium, gadolinium, and holmium. Across the design trajectory, the model shows monotonic improvement in neodymium selectivity as pocket offset and hydration gating are strengthened. The top candidate C5 reaches a neodymium selectivity gap of 0.4152 and a neodymium ground-state weight of 0.9856, while the order-four QCM estimate converges to 0.3418 against an exact reduced-model ground energy of 0.342. These results support a data-efficient and interpretable route for quantum-guided bioseparation design.
Download

Paper Nr: 79
Title:

Quantum Kernel Support Vector Machine for Quantum Dot State Recognition

Authors:

Md Rahatul Ashakin, Rubayat Khan, Avik Mahata, Mazharul Karim, Nuruzzaman Sojib and Don Roosan

Abstract: Semiconductor quantum dot platforms require rapid recognition of charge states during automated device tuning, especially when labeled data are scarce and device-to-device variation is strong. This study evaluated state recognition from 100 by 100 two-gate current maps using 30 by 30 labeled patches under a strict device-level split. Patch features were transformed with a signed logarithmic scale, standardized, compressed to four principal components, and scaled to the interval from zero to two pi. A fidelity quantum kernel support vector machine implemented in IBM Qiskit with a four-qubit ZZFeatureMap was compared against linear and radial basis function support vector machines across few-shot budgets of 10, 20, 40, and 80 samples per class. At 80 samples per class, the quantum kernel model achieved accuracy and macro F1 near 0.947 on held-out devices, outperforming the classical baselines in the evaluated setting. Noiseinjection experiments showed stable macro F1 under perturbation. These findings support fidelity-based quantum kernels as practical components for automated quantum dot tuning pipelines requiring few-shot generalization.
Download

Paper Nr: 82
Title:

Scored Rule Sets for Interpretable Multiclass Classification

Authors:

Robin Nunkesser

Abstract: Interpretable multiclass classification is studied across several research communities, but comparisons and conceptual integrations across rule sets, rule lists, decision trees, and evolutionary approaches are still limited. This paper addresses this gap by proposing scored rule sets as a unifying and interpretable hypothesis space. We first analyze structural relationships among established algorithms and identify indications of research siloing, especially between evolutionary and non-evolutionary lines of work. We then compare selected highly interpretable multiclass learners on UCI benchmark datasets. Beyond empirical comparison, we provide a formal definition of scored rule sets, show how they generalize classical rule-based models, and illustrate concrete transformations from logicGP, ExSTraCS, and CART representations. The resulting perspective clarifies when scored rule sets preserve interpretability while increasing modeling flexibility. Overall, the results support scored rule sets as a practically useful bridge between existing interpretable model families and as a basis for future cross-community method development.
Download

Paper Nr: 87
Title:

Enhancing Wildfire Prevention: An Operational Pipeline Integrating WRF Mesoscale Modelling and Automated CFFDRS Calculation

Authors:

Afonso Oliveira, Eva Maia and Isabel Praça

Abstract: Across Europe, climate change and shifting land-use patterns are driving longer fire seasons and increasingly extreme wildfire events, increasing the need for predictive tools to support proactive fire management. While traditional fire danger indices relied on observations from sparse meteorological stations, modern approaches use numerical weather prediction models that provide continuous spatial coverage and improved early warning capabilities. However, an important operational gap remains: although global forecast models offer broad coverage, their coarse resolution limits their usefulness for smaller geographical domains, where localized, terrain-driven microclimates strongly influence fire behaviour. To address this, this paper presents an automated computational framework that connects large-scale atmospheric data with localized fire danger calculations. The framework dynamically downscales global forecast data using the Weather Research and Forecasting (WRF) mesoscale model, generating high-resolution meteorological grids suitable for fire danger assessment. To evaluate operational reliability, WRF outputs were compared with ERA5-Land data using Pearson correlation, mean bias, and root mean square error (RMSE). A case study also demonstrates the automated calculation of an empirical fire index and its components, including the Fine Fuel Moisture Code (FFMC) and Buildup Index (BUI), at hourly resolution using open-source Python tools. Results show the framework supports scalable, localized wildfire risk assessment.
Download

Paper Nr: 95
Title:

Uncertainty Quantification for Annual Budgets: Bootstrap Methods on Short Monthly Time Series

Authors:

Emmanuel Zarpas and Martín Gómez-Abejón

Abstract: We evaluate five bootstrap methods within a calibrated ETS pipeline for constructing nominal 90% prediction intervals on the annual sum of short monthly time series (n = 36), using 49 428 series from the M3 and M4 benchmarks. Standard parametric intervals achieve only 76–78% coverage. Three findings emerge: (i) posthoc dispersion scaling (κ) improves coverage by 7–9 percentage points uniformly across methods; (ii) no bootstrap alternative significantly outperforms the Gaussian baseline after κ-correction; and (iii) empirical methods preserve residual skewness through summation while Gaussian methods do not—a distinction relevant for Monte Carlo scenario analysis but not captured by the Interval Score. The κ values are consistent across both benchmarks, supporting fixed defaults for short monthly series of similar provenance.
Download

Paper Nr: 98
Title:

Advancing m6Am Site Prediction Through Deep Learning on Diverse mRNA Sequence Landscapes: The m6Am-DLcat Approach

Authors:

Alfredo Cuzzocrea, Tasmin Karim, Md. Shazzad Hossain Shaon, Md. Fahim Sultan and Mst Shapna Akter

Abstract: Accurate identification of m6Am RNA methylation sites is crucial for biomedical research due to its link to diseases like cancer. Here, we introduce m6Am-DLcat, a novel CNN-Attention Deep Learning framework for m6Am site prediction in mRNA sequences. The model integrates five complementary feature descriptors (NCP, ASDC, PseEIIP, DBE, and K-mer), followed by a tree-based refinement step via LightGBM, XGBoost, and Chi-square. Evaluated against multiple deep learning and machine learning baselines, m6Am-DLcat outperformed state-of-the-art predictors, achieving a peak accuracy of 87.76% using XGBoost-selected features.
Download

Paper Nr: 100
Title:

A Two-Stage Clustering Framework for Profiling E-Commerce Search Queries

Authors:

Melis Öztürk Umut, Enis Teper and Mustafa Keskin

Abstract: In e-commerce platforms, predicting the correct product category for a search query is essential for delivering relevant results. However, traditional evaluation methods typically rely on broad aggregate metrics that lack the granularity to capture different variations in prediction quality. To address this, we propose a two-stage diagnostic framework that groups similar searches to pinpoint where performance varies. In the first stage, we segment queries based on their text structure, which allows us to focus specifically on brand-based clusters. In the second stage, we analyze the variance and characteristics of the retrieved products, focusing on the attributes, pricing, and categorical diversity of the search engine's output for each query. This dual-layer approach bridges the gap between simple text analysis and the actual search result landscape, allowing us to infer distinct shopping intents such as "Focused Buyers" targeting specific product sets and "Confused Browsers" associated with high category dispersion. Although this study primarily focuses on behavioral sub-clustering within brand-related query segments, the developed two-stage framework is designed to be inherently generalizable. By isolating these specific points, our framework provides search engineers with a clear roadmap for targeted interventions, such as improved brand clarification and smarter query rewriting, ultimately creating a more intuitive and effective search ecosystem.
Download

Paper Nr: 101
Title:

Hierarchical Structure-Based Influence Maximization

Authors:

Longyun Wang, Qing Zhang, Min Wu, Huijun Zheng, Wenchuan Yang and Xin Lu

Abstract: Influence maximization is a fundamental problem in complex networks, yet existing methods often suffer from high computational cost and stochastic instability. In this paper, we propose HSIM, a hierarchical structure-based influence maximization algorithm that enables efficient and deterministic evaluation of node influence. By extending hierarchical structures from individual nodes to node sets, we introduce the Hierarchical Structure Influence (HSI) metric to estimate infection scale without relying on extensive Monte Carlo simulations. HSIM further improves efficiency through peripheral node pruning and a lazy-forward selection strategy. EEx-perimental results on both synthetic and real-world networks demonstrate that HSI closely matches simulation-based evaluations and achieves correlation coefficients mostly above 0.97 with Monte Carlo ground truth, while maintaining high stability. Compared with state-of-the-art algorithms such as IMM, HSIM achieves comparable influence spread but improves computational efficiency by up to two orders of magnitude. These results indicate that HSIM is a scalable and robust solution for large-scale influence maximization tasks. Potential extensions include adapting the framework to dynamic, weighted networks and competitive diffusion scenarios.
Download

Paper Nr: 104
Title:

Outlier Detection in Low-Cost Air Quality Sensor Networks: Distinguishing Sensor Errors from Environmental Events

Authors:

Victoria Beltran, Daniel Fernandez-Garcia and Jose Santa

Abstract: Low-Cost air quality monitoring stations enable dense urban deployments but are prone to sensor malfunctions that generate erroneous data. A key challenge is distinguishing these sensor-induced anomalies from legitimate pollution episodes. This paper analyzes hourly CO concentration data from 23 low-cost stations in Cartagena (Spain) and evaluates eight outlier detection methods spanning classical statistical approaches and temporally-aware machine learning methods. To quantitatively assess each method’s ability to detect sensor errors without flagging environmental events, we propose a Monte Carlo validation framework that injects synthetic anomalies (spikes, offsets, dropouts) alongside simulated environmental episodes into real station data. Results show a fundamental recall versus false detection trade-off: no method exceeds 52% sensor error recall without also flagging over 8% of environmental events. Offset detection emerges as the critical weakness across all methods, with even the best detectors recovering barely 27% of sustained calibration drifts.
Download

Paper Nr: 109
Title:

A Statistical Analysis of the State-of-the-Art Transformer-Based Remaining Useful Life Models on C-MAPSS FD001

Authors:

Gabriel Vinicius Boin Freitas and Yuri Kaszubowski Lopes

Abstract: Transformer-based architectures have become dominant in Remaining Useful Life (RUL) prediction on the NASA C-MAPSS FD001 benchmark, where incremental improvements are frequently reported as new stateof-the-art results. However, such claims are often presented as single-point Root Mean Square Error (RMSE) values, with limited statistical contextualization or reproducibility assessment. This paper investigates whether recent performance gains on FD001 reflect robust methodological progress or statistical concentration within a stabilized benchmarking ecosystem. We combine: (i) a focused systematic review of 21 high-performing Transformer-based and Transformer-related models, (ii) an interquartile range (IQR) analysis of reported FD001 RMSE values, and (iii) a controlled 30-run reproducibility study of AGATT. The results reveal strong convergence in architectural hybridization patterns, hyperparameter choices, and evaluation protocols, together with a compressed reported RMSE distribution. The reproduced AGATT distribution, with RMSE = 15.99 ± 1.39, remained above the originally reported single value of 11.45, indicating that this value was not recovered as the central tendency under repeated execution. These findings do not prove universal architectural saturation across the RUL literature, but they provide evidence that FD001 has entered a regime in which marginal RMSE differences should be interpreted with variance-aware and reproducibility-oriented criteria.
Download

Paper Nr: 115
Title:

Neural Network-Based Group Sparsity for Nonlinear Feature Selection in Multi-Output Problems

Authors:

Mamadou Kanouté and Florence Forbes

Abstract: In high-dimensional settings, building predictive models that achieve strong generalization while maintaining small complexity requires effective dimensionality reduction. Eliminating redundant and irrelevant variables is therefore a crucial step in model design. In this paper, we propose a nonlinear feature selection method based on a single-hidden-layer neural network that incorporates regularization to jointly enforce group-wise sparsity and within-group variable selection. The proposed framework identifies relevant groups and informative variables within them, and can be applied to both multi-output regression and multiclass classification. The effectiveness of the approach is demonstrated on synthetic and real-world datasets. Results show that it successfully removes irrelevant groups while performing refined within-group selection, leading to improved interpretability and competitive predictive performance with minimal information loss.
Download

Paper Nr: 120
Title:

Age- and Sex-Specific Normative Profiles of Reactive Strength Index in Football Athletes: A GAMLSS Approach

Authors:

P. D. Chougale and U. Ananthakumar

Abstract: The Reactive Strength Index (RSI), defined as the ratio of flight time to contact time during drop jumps, is a critical metric for assessing stretch-shortening cycle efficiency in football athletes. Despite its widespread use, sex-specific age-related normative reference values remain underdeveloped. This study employed Generalized Additive Models for Location, Scale, and Shape (GAMLSS) with the Box-Cox Cole and Green (BCCG) distribution to develop comprehensive centile curves for RSI in male and female football players aged 8–32 years. Data were collected using ForceDecks force plate technology across both limbs. Results revealed distinct age-related patterns between sexes, with males demonstrating progressive increases in RSI across all ages, while females exhibited peak performance during late adolescence followed by gradual decline. The 50th percentile for males increased from 0.77 at age 8 to 1.84 at age 32, whereas females showed a decrease from 1.18 at age 8 to 0.82 at age 32 after reaching a peak during adolescence. These findings provide evidence-based benchmarks for talent identification, training prescription, and return-to-play decision-making in football, while highlighting the necessity of sex-specific assessment protocols.
Download

Paper Nr: 129
Title:

Ontology-Driven Semantic ETL for Executable Knowledge Engineering

Authors:

Antony Seabra, Daniel Schwabe and Sergio Lifschitz

Abstract: Extract–Transform–Load (ETL) pipelines play a central role in integrating heterogeneous data sources for analytics and decision support. However, traditional ETL pipelines struggle to capture complex semantic relationships in unstructured data, while Large Language Models (LLMs), despite their extraction capabilities, lack logical rigor and are prone to hallucinations. This paper introduces an Ontology-Driven Semantic ETL, a hybrid semantic framework for transforming heterogeneous data into formal, executable knowledge models. The methodology separates informational structuring, semantic grounding, and knowledge materialization across the Extract, Transform, and Load layers, decoupling knowledge engineering into TBox induction and ABox population. It integrates LLM-based extraction with ontology alignment to produce logic-based models implemented in Prolog, enabling deterministic reasoning over domain knowledge. The approach is evaluated using real-world health-domain documents and compared with LLM-based knowledge graph and retrieval-based baselines. Results show that our approach supports complex deductive queries not expressible in conventional BI or vector-based systems, while improving precision and recall in entity and relation extraction. The resulting models provide greater completeness, consistency, and explainability through explicit semantic grounding. These findings highlight the importance of integrating ontology-based knowledge construction into ETL pipelines, bridging data engineering and knowledge representation, and enabling a verifiable alternative to standard RAG-based architectures.

Paper Nr: 130
Title:

AgentSearch: Learning Efficient Agentic Workflows via Deliver Tree Search

Authors:

Dinesh Chowdary Attota and Ying Xie

Abstract: Multi-Agent systems powered by Large Language Models have demonstrated remarkable capabilities; however, their effectiveness is limited by rigid, manually designed workflows that do not adapt to varying task complexity. Existing adaptive methods utilize greedy policies that select workflow components without anticipating downstream cost-accuracy tradeoffs, often necessitating multiple attempts to identify successful configurations. This work introduces AgentSearch, a cost-aware Monte Carlo Tree Search (MCTS) framework that constructs agentic workflows through deliberative lookahead search. The proposed approach employs a dual-network architecture: a value network that decomposes expected rewards into success probability and remaining cost estimates, and a grammar-constrained policy network that ensures semantically valid constructions. Guided by these networks, MCTS explicitly simulates candidate workflow trajectories, enabling principled reasoning about the cost-accuracy tradeoff prior to decision-making. The networks are trained using a multi-phase protocol that combines stochastic exploration, supervised pre-training, and self-play refinement. Experiments on mathematical reasoning (MATH, AQUA-RAT) and code generation (HumanEval, MBPP) benchmarks demonstrate that the method achieves 80.03% average accuracy, surpassing Chain-of-Thought by 13.78 points and the best multi-agent baseline by 8.87 points. Notably, the approach attains single-episode success while reducing computational costs by up to 47%, thereby eliminating the trial-and-error exploration required by previous adaptive methods.
Download

Paper Nr: 137
Title:

Data First, AI Second: A Four-Lenses Conceptual Architecture for Enabling AI to Deliver Precision Impact

Authors:

Christopher J. Garasi and Aaron J. Moreno

Abstract: The promise of generative AI remains largely unfulfilled in high-consequence scientific and engineering organizations, not because AI capability is insufficient, but because the data infrastructure required to focus that capability into trustworthy, mission-relevant outputs has not kept pace with the capability itself. This position paper argues that AI-readiness is fundamentally a data management problem and introduces the Four-Lenses conceptual architecture: a framework for the data infrastructure conditions under which generalized AI capability can deliver precision impact in high-consequence contexts. The four lenses (Digital Representation, Knowledge Systems, AI-Ready Knowledge, and Validation) form a strict dependency chain rather than a checklist. Each maps to an established counterpart in Modeling and Simulation Verification, Validation, and Uncertainty Quantification (VVUQ), translating that mature epistemic standard into data infrastructure requirements for AI deployment; both address the same epistemic problem, in which an output's relationship to physical reality depends on input quality, model validity, and the appropriateness of methods for the specific application and risk tier. The architecture is explicitly a conceptual framework and research agenda. The paper assesses where current approaches (including RAG, GraphRAG, and emerging uncertainty quantification methods) satisfy these necessary conditions, where they do not, and what remains to be built.
Download

Paper Nr: 143
Title:

DeepSeekMath Meets Order Book: Group-Aware Policy Optimization for High-Frequency Directional Trading

Authors:

Sayak Chakrabarty and Souradip Pal

Abstract: This paper studies reinforcement learning for high-frequency trading on limit order books by pairing an Order-Flow-based state model with policy-gradient methods. Instead of value-based RL techniques like tabular Q-learning, our approach deploys policy-based methods like vanilla PPO and DeepSeekMath-inspired variants like GRPO and GSPO, that use group-normalized updates and downside-aware shaping. On backtests with financial assets AMZN, AAPL, and GOOG under a simplified backtesting setup based on spread-scaled re-wards, these new policies improve net average PnL, profitability, and drawdown over the Q-Learning baseline. Our results show that (1) Order-Flow signals are an adequate state for policy RL and (2) group-aware PPO surrogates are preferable over value-based baselines.
Download

Paper Nr: 144
Title:

Intent-Aware Predictive Orchestration for Softwarized Networks Using LLM-Assisted Policy Translation

Authors:

Pınar Ersoy and Mustafa Erşahin

Abstract: Softwarized networks have significantly improved programmability, elasticity, and service agility; however, network operations still depend heavily on expert-driven manual policy specification, particularly when translating high-level service intent into machine-executable orchestration actions. This gap becomes more pronounced in dynamic environments where traffic demand, resource pressure, and service-level objective violations evolve faster than rule-based operational workflows can respond. This paper presents an intent-aware predictive orchestration framework in which a large language model (LLM) serves as an intent-to-policy translation interface, while a predictive control layer estimates near-future service degradation risks and proactively selects orchestration actions. The proposed system consists of five coordinated modules: an intent input layer, an LLM-based intent parser, a policy validator with safety guardrails, a prediction engine for congestion and resource risk, and an orchestration layer for scaling, rerouting, and placement adaptation. Unlike works that employ generative models solely for textual assistance, the present study formulates the LLM as a constrained semantic compiler that maps operator intent to a structured network policy. The predictive layer jointly addresses three operational tasks: service-level agreement (SLA) violation prediction, link congestion prediction, and virtualized network function (VNF) saturation prediction. A benchmark-oriented evaluation is conducted against rule-based intent mapping, sequence labeling baselines, static threshold orchestration, and non-predictive reactive control. The proposed method achieves higher intent parsing accuracy, improved policy validity, lower SLA violation rate, and faster time-to-action than competitive baselines, while incurring bounded token and compute overhead. The results indicate that the practical value of LLMs in network softwarization is greatest when embedded in verifiable control loops rather than used as unconstrained generators.
Download

Paper Nr: 148
Title:

Chunking Is Not the Bottleneck: Why RAG Evaluation Must Look beyond Segmentation Strategy

Authors:

Aaditya Chauhan and Nivas Hegde

Abstract: The RAG community devotes disproportionate attention to optimizing text chunking strategy. Through a factorial experiment crossing four chunking strategies, two embedding models, two chunk sizes, and two QA datasets (Natural Questions and MS-MARCO), we find that, in QA-style retrieval on two established datasets, choice of dataset (24–78% of variance) and embedding model (11–50%) together explain 67–89% of retrieval quality variance, while chunking strategy accounts for only 1–2% and is not statistically significant (p > 0.21). An end-to-end LLM-as-judge evaluation confirms that chunking differences do not propagate to generation quality in this QA setting (Kruskal-Wallis p > 0.88). We argue that, for QA-style retrieval on general-purpose natural-language corpora, chunking strategy is a low-variance factor, and that RAG benchmarks should report variance decomposition alongside absolute metrics so practitioners can allocate engineering effort to the factors that actually drive performance.
Download

Paper Nr: 156
Title:

Text-to-SQL with Large Language Models: Challenges Revisited and New Dimensions

Authors:

Luca Sala, Giovanni Sullutrone and Sonia Bergamaschi

Abstract: Building on a prior analysis that identified five key challenges for LLM-based Text-to-SQL (response time, scalability, hallucinations, dataset representativeness, and knowledge acquisition), this position paper argues that each original challenge has given rise to a new dimension: response-time and scalability pressures produced agentic approaches; the pursuit of accuracy on complex queries led to reasoning models; hallucination research exposed the deeper trustworthiness problem, including security; static-benchmark limitations motivated conversational evaluation; and scalability became an economic question, surfacing the open versus proprietary tension. The field has transformed: agentic architectures now surpass 81% execution accuracy on the BIRD test set, yet state-of-the-art systems remain no higher than 17% end-to-end success on multi-turn interactive benchmarks, and backdoor attacks succeed with less than 1% poisoned training data. We synthesize these concerns and outline a research agenda along three horizons, arguing for trustworthiness, interactivity, and economic sustainability as first-class concerns.
Download

Paper Nr: 177
Title:

Individual-Level Probabilistic Model Integrating Rule Representation and Logistic Curves

Authors:

Akari Oda, Yoichi Seki and Kaoru Shimada

Abstract: We propose a framework that enables probabilistic outputs by introducing logistic function expressions into interpretable rule representations, facilitating the analysis of attributes related to prediction probabilities and their variations for individual cases in an interpretable manner. We discover itemsets with statistically distinctive backgrounds (ItemSBs) through an evolutionary computation method and identify candidates for attributes related to changes in probability for each case. By incorporating a logistic curve into the consequent part of each rule, we achieve probabilistic outputs as continuous values. This extended rule representation enabled rule-based probability prediction at the individual case level. Furthermore, by introducing analyses that evaluate probabilities, their changes, and the changes in probability when input attributes are hypothetically modified, we enabled the interpretation of the relevance of attributes in individual cases.
Download

Paper Nr: 183
Title:

AI-Powered Early Detection of Child Malnutrition Using Dimensionless Anthropometric Ratios and Deep Learning

Authors:

Vaishnu Kanna and Subhrakanta Panda

Abstract: Child malnutrition affects over 149 million children globally, yet gold-standard screening tools remain unavailable in most low-resource settings. This paper investigates whether dimensionless body and facial proportion ratios extracted from standard smartphone photographs using Google MediaPipe can serve as proxies for clinical anthropometric measurements to screen children for malnutrition. Three complementary pipelines are evaluated on the AnthroVision dataset of 2,389 children for simultaneous multi-label classification of stunting, underweight, and wasting. The Baseline pipeline (five clinical measurements) achieves F1 Macro 0.829 and AUC-ROC 0.944 with LightGBM. The Image-Only pipeline (14 MediaPipe-derived dimensionless ratios) yields F1 Macro 0.391 and AUC-ROC 0.648, confirming limited but above-chance discriminative signal from camera features alone. The Combined pipeline (20 features) achieves F1 Macro 0.851 and AUC-ROC 0.922, demonstrating that image ratios carry complementary nutritional signal. Transfer learning experiments with ResNet50 and EfficientNet-B0 reveal catastrophic overfitting on the 1,495-image training set, with EfficientNet achieving AUC 0.400 only when using Focal Loss. SHAP and LIME explainability analysis confirms clinically interpretable feature hierarchies. Results establish that camera-derived ratio features, while insufficient as standalone classifiers, meaningfully augment clinical screening pipelines.

Paper Nr: 187
Title:

BLEO: A Binary Language Education Optimization for Feature Selection

Authors:

Beatriz Souza Sé Barros, Guilherme N. Marques, Douglas Rodrigues and João Paulo Papa

Abstract: Language Education Optimization (LEO) is a population-based metaheuristic inspired by foreign language acquisition, organized into teacher-guided instruction, mutual learning, and individual practice. This paper introduces BLEO, a binary variant of LEO tailored for wrapper-based feature selection, a supervised classification setting in which the task is to select a compact subset of informative features while preserving predictive performance. BLEO adapts LEO’s continuous search dynamics to the discrete domain via a probabilistic binarization process driven by a transfer function, where each candidate solution is encoded as a binary feature mask indicating whether each attribute is selected or discarded. Experiments on 17 heterogeneous benchmark datasets compared BLEO with six well-established metaheuristics, focusing on classification accuracy, F1-score, and the number of selected features. Results demonstrate that BLEO achieves competitive performance on 15 of the 17 datasets and consistently achieves significant dimensionality reduction. Although predictive performance decreased slightly on two datasets and execution time is longer in most cases, reducing features by approximately 50% yields a leaner representation. These findings underscore BLEO’s effectiveness as a robust alternative for feature selection in high-dimensional classification problems, particularly in scenarios where compact feature subsets, model simplicity, and predictive quality are prioritized over search speed.
Download

Paper Nr: 190
Title:

Automated & Real-Time Privacy Quantification for Microservices Architectures

Authors:

Catarina Silva, Bernardo Falé, João Paulo Barraca and Paulo Salvador

Abstract: The decentralized nature of microservice architectures introduces privacy challenges that exceed the capabilities of traditional static auditing. To address this, we present a real-time privacy quantification framework based on the Privacy-sensitive Data Categorization (PsDC) model. By combining Layer 7 Deep Packet Inspection (DPI) with Natural Language Processing (NLP), our methodology continuously inspects the semantic intent of network payloads. We introduce the Privacy Exposure Index (PEI), a dynamic risk metric that links the detection of sensitive entities directly with their operational context. We validated this approach in a Ku-bernetes environment using a custom ingestion engine and analyzer. Experimental results show high fidelity in identifying complex threats, including DNS-based data exfiltration, and successfully isolated high-risk service interactions with localized PEI scores of 5.32. Ultimately, this work establishes a foundation for a proactive DevPrivOps lifecycle, demonstrating that semantic-aware observability can replace manual privacy checks and act as the core decision engine for active privacy enforcement.
Download

Paper Nr: 194
Title:

Time Series Forecasting of Sports Ticket Sales with Venue-Independent Fan Segments

Authors:

Shun Takagi, Haru Terashima, Naoki Tamura, Kazuyuki Shoji, Kenta Urano, Takuro Yonezawa and Nobuo Kawaguchi

Abstract: Online ticket sales in sports events accumulate detailed data such as daily sales volumes and purchaser attributes, establishing a foundation for optimizing sales strategies. Optimizing ticket sales requires time series forecasting during the sales period; however, this has not been sufficiently explored. Sports ticket sales have a short sales period, and the limited observations available as input to forecasting models make existing methods prone to reduced accuracy. Furthermore, the recent surge in new stadium construction has increased demand for forecasts at venues lacking historical data. In such situations, leveraging data from existing venues becomes crucial, yet inter-venue differences make it difficult to extract common patterns from total sales alone. In this study, we propose an input representation that decomposes total sales into fan segments based on visit frequency. This enables extraction of purchasing patterns common across venues. In our experiments, we first compared existing time series forecasting methods and confirmed that LSTM combined with fan segment decomposition achieved the highest accuracy. Furthermore, through comparison experiments varying input representations and input lengths, fan segment decomposition consistently demonstrated higher accuracy than other decomposition methods.
Download

Paper Nr: 199
Title:

SwitchEmbed: Representation Learning for Arabic-English Code-Switched Text

Authors:

Mariam Rizkallah, Amani Ghonim, Ahmed Sherif and Caroline Sabty

Abstract: Multilingual speakers often alternate between languages within a conversation, a phenomenon known as code-switching. This is common in Arabic-speaking communities, where Arabic and English are frequently mixed in everyday communication. Although recent advances in natural language processing have been driven by pretrained multilingual language models, these models are largely trained on monolingual data and often struggle to capture the abrupt language transitions and cross-lingual semantic interactions that characterize code-switched text. This work investigates representation learning for Arabic-English code-switched text at multiple levels. At the word level, we employ a code-switch-aware masked language modeling objective that captures token-level language variation and switch points. At the sentence level, we adopt a contrastive learning framework with natural language inference supervision to encourage semantically consistent sentence embeddings across monolingual and code-switched variants. To support this objective, we introduce CS-SNLI, a large-scale Arabic-English code-switched natural language inference dataset. The resulting embeddings are evaluated on sentiment analysis and named entity recognition. On sentiment analysis, the word-level pipeline improves F1-score by 2.17 percentage points over mBERT and 2.27 percentage points over XLM-R, while the sentence-level pipeline improves F1-score by 1.78 and 1.28 percentage points, respectively. In contrast, named entity recognition shows only marginal, statistically insignificant gains.
Download

Paper Nr: 220
Title:

Benchmarking YOLOv8 and YOLOv11 for Meteor Echo Detection in BRAMS Radio Spectrograms

Authors:

Diogo Ramalho Fernandes, Stijn Calders, Olivier Parisot, Mickaël Stefas, Hervé Lamy and Katrien Kolenberg

Abstract: Automatic meteor detection in BRAMS (Belgian RAdio Meteor Stations) spectrograms is challenging due to weak echoes near the noise level and overlapping radio artifacts. This paper benchmarks You Only Look Once (YOLO) detectors, focusing on YOLOv8 and YOLOv11, for meteor echo detection, evaluated under identical conditions across four data preparation strategies: raw RGB, grayscale, custom threshold-based denoising, and SNR-filtered annotations. Results show that YOLOv11-x achieves the strongest detection performance. Threshold denoising reduces recall by removing weak signal details alongside noise, while Signal-to-Noise Ratio (SNR)-based filtering yields only marginal improvement, suggesting that low-SNR echoes still carry useful information. A qualitative analysis of Radio Frequency Region of Interest Convolutional Neural Network (RFROI-CNN) denoising is also included as a learned preprocessing direction.
Download

Paper Nr: 221
Title:

Evaluating and Comparing Optimized CNN and Transformer Architectures for EEG-Based Depression Detection

Authors:

Khaoula Ansari, Ibtissam Abnane and Ali Idri

Abstract: Major Depressive Disorder (MDD) is a leading cause of global disability, affecting over 280 million people worldwide. Clinical diagnosis still relies on subjective tools such as questionnaires and behavioral observations, which limit accuracy and reproducibility. Electroencephalography (EEG) has emerged as a promising non-invasive modality for objective MDD detection. This paper evaluates and compares three deep learning architectures: EEGNet, Deep4Net, and a Vision Transformer (EEG-ViT), for binary MDD classification from EEG signals. A unified preprocessing pipeline was applied to two heterogeneous public EEG datasets: the EEG Dataset for MDD and MODMA. Hyperparameter optimization was performed using Particle Swarm Optimization (PSO) for CNN models and Bayesian optimization via Optuna for the transformer. A subjectwise evaluation protocol was adopted to prevent data leakage. Performance was assessed at both window and subject levels. All three models achieved a subject-level Balanced Accuracy (BAcc) of 0.800 over the MDD dataset. EEGNet+PSO achieved the best performance (BAcc = 0.833) over MODMA, while EEGViT+Optuna ranked last (BAcc = 0.633). No single architecture consistently outperformed the others across both datasets. Dataset characteristics, particularly subject pool size, were the primary determinants of performance.
Download

Paper Nr: 226
Title:

Identifying Brand Advocates: A Neural Network Approach to Social Media Analysis

Authors:

R. E. Loke and O. Ran

Abstract: The exponential growth in unstructured data paired with the shifting dynamic in consumer-brand interactions, presents significant challenges and opportunities for businesses aiming to harness customer advocacy (CA) to enhance brand reputation. This paper explores the application of a hybrid XLNet-BiLSTM neural network model, to identify CA from social media comments. The research investigates neural networks' ability to analyse vast amounts of unstructured data-primarily social media interactions-to discern patterns that signify CA. This study particularly focuses on the smartphone industry, where online customer engagement plays a critical role in influencing brand reputation and purchase decisions. Through extensive literature review and empirical data analysis, this paper applies a methodological approach that combines the contextual understanding capabilities of XLNet with the sequential data processing strength of BiLSTM. The model's effectiveness is evaluated based on its ability to classify social media comments accurately. Results indicate that while the model achieves high recall, indicating effectiveness in identifying potential advocates, the precision metric suggests a need for further refinement to reduce the rate of false positives. This paper contributes to the field by demonstrating the practical applications of advanced machine learning techniques in real-world business scenarios and suggesting pathways for future research to refine and enhance the identification of CA on social media platforms.
Download

Paper Nr: 227
Title:

Media-Derived Epidemic Intelligence for Avian Influenza: A Multi-Granularity Evaluation of LLM-Augmented Event-Based Surveillance

Authors:

Bibek Paudyal, Karine Deschinkel, David Laiymani and Christophe Guyeux

Abstract: Zoonotic outbreaks cause large-scale animal mortality and economic loss, yet official tracking systems typically lag biological onset by approximately two weeks. To address this, we introduce an AI-augmented EventBased Surveillance system that extracts early outbreak signals from multilingual news media. Analysing over 85,000 articles across 28 European countries (2021–2026), we deployed a local Large Language Model (Qwen 2.5-14B-Instruct) to classify and structure epidemiological data. Validated against 400 manually annotated articles using an expanded multilingual evaluation, the model achieves a macro-average F1-score of 0.75 across six extraction fields (species identification improved to F1 = 0.62 with multilingual synonym matching). Media-derived event dates match the WAHIS biological onset date and precede official WAHIS notifications by 13–14 days. At the country level, we achieve a ROC-AUC of 0.629 for France and 0.700 for Germany. After applying Platt-scaling probability calibration to linear models, Germany’s district-level Logistic Regression reaches a ROC-AUC of 0.772, outperforming tree-based methods (0.642). LLM-extracted urgency scores improve Germany’s country-level prediction by +4.1 pp, while semantic features provide no consistent lift for France, reflecting a dependence on national media structure. In our two-country evaluation, spatial granularity, epidemic base rate, and epidemic phase appear to dominate EBS performance more than model choice.
Download

Paper Nr: 229
Title:

Beyond Accuracy: Comparative Explainability Analysis of Random Forest and XGBoost Using SHAP and LIME

Authors:

Danilo Medeiros Eler, Leonardo Menosse Ribeiro, Thales Vinicius de Brito Ue and Wilson Estecio Marcílio Junior

Abstract: The increasing use of machine learning models in decision support systems has expanded the need for methods capable of making these solutions more transparent and interpretable. However, complex models often exhibit behavior that is difficult to understand, making it difficult to identify the factors responsible for automated decisions. In this context, this work presents a comparative explainability analysis between Random Forest and XGBoost models using SHAP and LIME techniques for global and local interpretation of predictions. As a case study, the German Credit dataset was used for credit risk analysis. The adopted methodology involved data preprocessing, model training, evaluation through classification metrics, and interpretable analysis of correct, uncertain, and incorrect cases. The results demonstrated that, although the models presented similar predictive performance, relevant differences were observed in decision patterns and sensitivity to the analyzed variables. Thus, the study highlights the potential of Explainable Artificial Intelligence techniques to increase the transparency, interpretability, and reliability of machine learning models in sensitive decision-making scenarios, allowing specialists to understand the factors associated with predictions and use this information to support human decision-making.
Download

Paper Nr: 238
Title:

SENTINEL: Semantic Validation of Financial News Clustering with LLM-Based Evaluation

Authors:

Braeden Mefford, Andrew Berg, Houston Miles and Mia Y. Wang

Abstract: Financial news sentiment is increasingly used in downstream machine learning systems, yet the structural validity of headline-level sentiment signals remains underexplored. We present SENTINEL, a validation pipeline that evaluates whether financial news headlines form semantically coherent topic clusters and whether sentiment labels remain consistent within those clusters. Using 124,330 headlines collected from public financial news sources, we generate sentence embeddings, perform dimensionality reduction and clustering, and evaluate cluster quality using an LLM-based scoring framework. We additionally measure clustering robustness across independently seeded runs to assess stability. Experimental results show strong topic coherence, sentiment consistency, and cross-seed stability, indicating that financial news headlines exhibit sufficiently stable semantic structure to support downstream ML applications.
Download

Paper Nr: 239
Title:

Evaluating Retrieval-Augmented Generation on Social Bias Benchmarks across Small Language Models

Authors:

Manuel Amir Freer Valdez and Arghir-Nicolae Moldovan

Abstract: As small-scale, open-source Large Language Models (LLMs) proliferate for on-device and privacy-centric applications, understanding the trade-offs between their utility and behavioural reliability becomes critical. This study evaluates a suite of instruction-tuned LLMs, Gemma, Llama, and Qwen (≤4B parameters), treating the model family, and scale as the primary units of analysis. Retrieval-Augmented Generation (RAG) is employed as a controlled experimental condition to assess utility gains on the Natural Questions (NQ) benchmark, while utilizing native, non-augmented configurations to establish a fairness baseline via the Bias Benchmark for QA (BBQ). The findings reveal that while RAG significantly enhances utility, often doubling Exact Match (EM) scores, these gains are non-uniform and architecture-dependent, with certain families exhibiting greater "retrieval-readiness" than others. Paradoxically, the fairness analysis shows that providing explicit context in disambiguated settings can increase stereotype engagement rather than suppressing it. These results suggest a fundamental disconnect between a model's capacity for factual accuracy and its ability to maintain social fairness, highlighting the need for multi-dimensional evaluation frameworks for small-scale systems.
Download

Paper Nr: 41
Title:

Towards Realistic Privacy-Preserving Stress Detection for Nurses Using Smartwatch Data

Authors:

Cedric Aebi, Albin Grataloup, Souhir Ben Souissi and Christoph Golz

Abstract: Recent advancements in wearable technologies, such as smartwatches, have enabled continuous, non-invasive health monitoring. The data collected from these devices can be processed using various machine learning models to predict different aspects of an individual’s health. Chronic stress is particularly concerning due to its long-term health implications, making the prediction of stress levels crucial. Typically, machine learning models require substantial amounts of data to achieve high performance, which raises significant privacy concerns, as individuals may be unwilling to share their private and sensitive information. In this study, we investigate privacy-preserving machine learning approaches for stress detection, including individual learning and federated learning, where models are trained locally on each client’s data. Our analysis explores the trade-offs between generalization and personalization, as well as the privacy implications. We adopt a realistic scenario using smartwatch data from nurses, treating each participant as a separate client. Additionally, we assess the utility of LiteRT for enabling efficient inference after the training processes on resource-constrained devices. Our results indicate that the global model produced by federated learning demonstrates limited performance, whereas individual learning achieves F1 scores as high as 0.99.
Download

Paper Nr: 42
Title:

Behavioral Profiling of Active E-Commerce Customers: Integrating Post-Purchase Interaction Signals into Comparative Clustering Models

Authors:

Serena Tomakyan, Arda Raştak and Nilay İşeri

Abstract: Customer experience (CX) has become a critical differentiator in competitive e-commerce environments characterized by heterogeneous customer behavior. While survey-based and loyalty-oriented metrics are widely used to monitor CX, they often fail to capture operational inefficiencies that directly affect service complexity and cost-to-serve. This study presents a comparative clustering framework for behavioral profiling of active e-commerce customers by integrating multi-temporal engagement, economic value, and issue-based operational metrics. Four clustering paradigms, namely K-Means, Agglomerative Hierarchical Clustering, Gaussian Mixture Models, and a Self-Organizing Map-based hybrid approach, are evaluated under identical conditions. Based on internal validation indices, K-Means provides the most stable and interpretable segmentation structure for the analyzed dataset. The resulting four-cluster solution reveals distinct behavioral profiles that differ in purchase intensity, revenue contribution, and operational inefficiencies. The findings demonstrate that issue-aware behavioral segmentation can support actionable customer experience management and resource optimization strategies.
Download

Paper Nr: 60
Title:

Evaluating Nearest-Neighbor Variants for Time Series Classification with Manhattan-Based Measures

Authors:

Zoltán Gellér, Vladimir Kurbalija and Mirjana Ivanović

Abstract: Due to its simplicity of implementation and ease of understanding, the nearest-neighbor (NN) method is among the most popular classifiers for both practical applications and benchmarking, with the choice of distance (similarity) measure representing a critical factor. In this paper, we compare the performance of the 1NN classifier with majority-voting and weighted kNN variants, employing several Manhattan-based distance measures. The results demonstrate that all evaluated kNN variants outperform 1NN in average accuracy across the 128 UCR time series datasets, with 1NN exhibiting statistically inferior performance primarily in comparison to the following weighting functions: Dudani, dual-distance, and dual-uniform.
Download

Paper Nr: 63
Title:

Sequence-Constrained Data Augmentation: A First- and Second- Order Markov Transition Matrix Approach

Authors:

Hanyu Li, Yuanbo Kong and Meng Wang

Abstract: Move transition sequence constraints serve as the core mechanism for characterizing the prior rhetorical structure of academic discourse. They are fundamentally important for ensuring the sequential self-consistency of augmented data and enhancing the functional completeness of minority-class move samples. Existing data augmentation methods typically treat individual sentences as independent operational units, completely ignoring inter-move transition patterns. Consequently, augmented samples often suffer from functional overlap and semantic fragmentation at the sequence level. To address these issues, this paper proposes a novel data augmentation approach that embeds sequence constraints via first- and second-order Markov transition matrices. Specifically, we construct a move sequence constraint mechanism that systematically calculates the co-occurrence frequencies of adjacent move pairs and ternary move combinations from annotated corpora. To resolve the zero-probability estimation problem for low-frequency triplets, we apply additive smoothing fused with linear interpolation. We then employ a constrained random walk strategy to batch-sample structurally valid move sequence skeletons. This sampling is governed by three constraints: transition probability thresholds, reasonable sequence length intervals, and stylistic norms for starting and ending moves. The sampled skeletons are subsequently filtered based on sequence validity scores, thereby elevating the granularity of the augmentation operation from the sentence level to the sequence level. Driven by these structurally valid move sequence skeletons as global conditions, we prompt Large Language Models (LLMs) to generate augmented text. This process implicitly encodes discourse position-aware information into the augmented samples, effectively minimizing the occurrence space of inter-class functional overlap. Experimental results demonstrate that our proposed method achieves consistent performance improvements across all evaluated base LLMs and encoder configurations. Under the optimal configuration (Qwen3-8B + DeBERTa-v3-base), the Macro-F1 score reaches 80.84%, yielding an absolute improvement of 14.23% over the unaugmented baseline. For minority classes, the F1 score for the Gap category improves from 50.34% to 74.89% (+24.55%), while the Limitation category rises from 42.78% to 71.34% (+28.56%). The cumulative gain across these two minority classes is approximately double the overall Macro-F1 gain. Furthermore, ablation studies verify that removing the sequence constraints causes the performance of the Gap and Limitation categories to drop significantly by 9.02% and 9.89%, respectively—nearly twice the overall drop in Macro-F1. This confirms the irreplaceable role of the sequence constraint mechanism in directionally repairing the decision boundaries of minority classes. Ultimately, the proposed method provides a new technical pathway for resolving class imbalance in low-resource academic discourse analysis. It exhibits broad application prospects in areas such as academic move recognition, automatic rhetorical structure analysis, and intelligent academic text annotation.
Download

Paper Nr: 78
Title:

Quantum Pharmacogenomics for Polypharmacy Risk Minimization in Elderly Care

Authors:

Don Roosan, Mazharul Karim, Rubayat Khan and Inyene Essien-Aleksi

Abstract: Pharmacogenomics can reduce adverse drug reactions by tailoring therapy to genetic variation, but older adults with polypharmacy require decision support over many interacting drug combinations. This paper presents a quantum pharmacogenomics pipeline that integrates curated PharmGKB gene-drug evidence, DrugBank interaction structure, and proxy electronic health record complexity signals, then formulates regimen selection as a QUBO solved with QAOA in IBM Qiskit. A pilot cohort of 300 elderly polypharmacy cases with 12 candidate drugs, 48 interaction edges, and 10 pharmacogenomic genes was evaluated against a classical greedy baseline and an oracle optimum. QAOA reduced mean predicted adverse drug reaction risk from 7.148 to 6.795 and lowered the high-risk rate above 7.023 from 0.49 to 0.36. It improved over greedy selection in 0.59 of cases and matched the oracle optimum in 0.997 of cases. Energy-landscape visualization showed probability mass concentrated on low-risk regimens while preserving near-optimal alternatives. These results support quantum optimization as an interpretable decision layer for precision polypharmacy management. Because the pilot cohort is generated in-script and ADR risk is estimated from a constructed objective rather than observed longitudinal outcomes, the results should be interpreted as a controlled proof-of-concept rather than clinical validation.
Download

Paper Nr: 81
Title:

From Prediction to Decision: A Counterfactual Machine Reasoning Framework for ESG Analysis

Authors:

Tanzina Nizam and Yeon Kyupil

Abstract: Environmental, Social, and Governance (ESG) evaluation is traditionally treated as a predictive task, where machine learning models estimate scores from financial and contextual features. Such approaches remain fundamentally limited: they provide predictions without structured reasoning, fail to resolve conflicting signals, and cannot support counterfactual decision analysis. This paper proposes a Machine Reasoning (MR) framework that transforms ESG evaluation into a structured decision-making process. The system decomposes ESG evidence into three independent streams: environmental efficiency, financial comparative position, and causal profit-margin effects estimated via DoWhy, and integrates them through five conditional reasoning regimes that resolve conflicts rather than average them. The architecture possesses three properties absent from standard ML pipelines, explanations are produced by the same conditional logic that generates predictions, not inferred post-hoc; hard weight discontinuities at regime boundaries prevent financial strength from compensating for environmental failure; and counterfactual interventions re-run the full reasoning pipeline, capturing non-linear regime shifts that surrogate-model approaches cannot represent. Validated on 11,000 firm-year observations without lagged ESG inputs, the fusion model achieves R²=0.641, a +0.44 R² gain over financial-only baselines with structured decision traces and intervention analysis as additional outputs.
Download

Paper Nr: 91
Title:

A Multi-Mode Online Learning Framework for Drift-Aware Maritime Fuel Consumption Prediction

Authors:

Nadeem Iftikhar and Finn Ebertsen Nordbjerg

Abstract: Accurate fuel consumption prediction is important for maritime efficiency and sustainability. However, maritime operations are inherently non-stationary, and models trained on static datasets may degrade after deployment as operating conditions change. This paper presents a drift-aware Online Machine Learning (OML) framework with four complementary adaptation mechanisms: (i) incremental learning for gradual change, (ii) rolling model refresh for sustained shifts, (iii) real-time model switching for abrupt degradation, and (iv) predictive updating to prepare for forecasted disturbances. The controller selects among these modes using drift and model-health signals, with the aim of maintaining predictive accuracy while reducing unnecessary retraining. The paper describes the overall architecture, monitoring design, and evaluation protocol using real vessel telemetry with exogenous weather covariates. A sanitized synthetic replay is also used to illustrate all mode triggers under controlled conditions. The results show improved performance over static baselines and more targeted adaptation under changing maritime conditions.
Download

Paper Nr: 93
Title:

Integrating Causal Inference with Graph Neural Networks for Alzheimer’s Disease Analysis

Authors:

Pranay Kumar Peddi and Dhrubajyoti Ghosh

Abstract: Deep graph learning has advanced Alzheimer’s disease (AD) classification from MRI, but most models remain correlational, confounding demographic and genetic factors with disease-specific features. We present Causal-GCN, an interventional graph convolutional framework that uses do-operator-inspired perturbations with covariate adjustment to identify brain regions with stable model-based influence on AD probability. Each subject’s MRI is represented as a structural connectome where nodes denote cortical and subcortical regions and edges encode anatomical connectivity. Confounders such as age, sex, and APOE4 genotype are summarized via principal components and included in the causal adjustment set. After training, interventions on individual regions are simulated by severing their incoming edges and altering node features to estimate average causal effects on disease probability. Applied to 484 subjects from the ADNI cohort, Causal-GCN achieves performance comparable to baseline GNNs while providing interpretable causal effect rankings that highlight posterior, cingulate, and insular hubs consistent with established AD neuropathology.
Download

Paper Nr: 121
Title:

Agreement beyond Chance in LLM-Based Hazard Severity Classification for Railway Functional Safety

Authors:

Padma Iyenghar

Abstract: Severity classification of hazards is a critical step in railway functional safety processes under the CENELEC EN 50126/50128/50129 framework. This paper presents a structured empirical evaluation of agreement beyond chance between LLM-generated severity predictions and expert-labeled classifications using a regulator-aligned benchmark dataset (TeSiP Severity Benchmark) derived from the TeSiP/SIRF function catalog and released as an open dataset. Under fixed prompting and deterministic decoding, higher-capacity proprietary LLMs achieve statistically significant agreement with expert severity labels beyond chance and demonstrate higher consistency than the evaluated open-weight models, although overall agreement remains moderate. Confusion analysis reveals systematic class-distribution bias and persistent ambiguity at adjacent severity boundaries. These results indicate that deterministic LLM-based classification can approximate expert-aligned severity assignments to a limited but measurable extent, while remaining sensitive to boundary ambiguities inherent in normative classification. The findings provide quantitative evidence on the capabilities and limitations of LLMs for regulator-aligned severity classification, supporting their use as structured decision-support tools within supervised railway functional safety workflows.
Download

Paper Nr: 136
Title:

Developing an ML Model for Orthophoto Segmentation Using Automatically Generated Labels

Authors:

Kevin Kocon and Michel Krämer

Abstract: In this paper, we develop a machine learning (ML) model for the segmentation of orthophotos in urban areas throughout Germany. We present our ML pipeline as well as the results of evaluating the model. The main challenge is that labels required for training do not exist for the classes we want to identify in the available aerial imagery. Manual labelling is too time-consuming and the commonly used approach of generating synthetic training data has the disadvantage that artificial data cannot fully represent real-world characteristics. Instead, we generate our labels automatically based on official cadastral information, as well as results from an existing convolutional neural network (CNN) segmenting 3D point clouds from car-based terrestrial mobile mapping. This is a novel approach that, to the best of our knowledge, has not been explored before. Our evaluation shows that the model reliably identifies all classes throughout Germany, regardless of region, season, or camera. Notably, the model correctly classifies surfaces even in areas where the training labels contained no information. The predictions are thus visually closer to reality than the generated labels themselves. Our approach therefore represents a viable alternative in cases where there are no labels available for training. The resulting model is deployed in production by Deutsche Telekom for fiber Internet rollout planning.
Download

Paper Nr: 138
Title:

CITRUS: Screening the Past, Ranking the Future

Authors:

Sayak Chakrabarty, Souradip Pal and Imon Banerjee

Abstract: Sequential recommender systems increasingly rely on attention-based architectures to model dependencies in user histories, but dense candidate-to-history attention can still overemphasize irrelevant interactions and reduce robustness in dynamic recommendation settings. In this position paper, we present CITRUS: Continuous-Time Influence Recommendation Using Screening, a modular sequential recommendation architecture instantiated with three components: (i) nonlinear Granger-style parent selection using kernel-based screening, (ii) geodesic attention based on normalized cosine similarity on a spherical manifold, and (iii) continuous-time temporal influence through exponential lag decay. Unlike earlier architectures, the proposed model uses a deterministic top-k screening mask to select the top-k history items with the highest nonlinear dependence (measured by an RBF kernel / HSIC proxy) with each candidate item, enabling sparse and nonlinear causal attention. We evaluate the model on three public benchmarks covering news and e-commerce recommendation: MIND, Globo, and Amazon. Across these datasets, CITRUS achieves 0.4903 nDCG@10 / 0.4319 MRR@10 on MIND, 0.5953 nDCG@10 / 0.5938 MRR@10 on Globo, and 0.5376 nDCG@10 / 0.5053 MRR@10 on Amazon, demonstrating that the proposed sparse-geometry-temporal combination remains effective across heterogeneous sequential recommendation settings.
Download

Paper Nr: 146
Title:

Progressive Multi-Objective Optimization for Improved t-SNE Embeddings

Authors:

Samir Brahim Belhaouari, Skander Bensegueni, Lyes Fennour and Dhia El Hak Amani

Abstract: Dimensionality reduction is essential for analyzing and visualizing high-dimensional data, with t-distributed Stochastic Neighbor Embedding (t-SNE) being widely used due to its ability to preserve local neighborhood structures. However, its reliance on a single Kullback–Leibler (KL) divergence objective often leads to poor global structure preservation and sensitivity to local inconsistencies. In this paper, we propose a progressive multi-objective optimization framework that enhances t-SNE by integrating complementary loss functions, including a ranking-aware divergence (KLmax) and a Wasserstein-based term for global alignment. Rather than optimizing all objectives simultaneously, we introduce a progressive training strategy that gradually incorporates these components, enabling more stable convergence and improved embedding quality. Additionally, the framework is applied to latent representations learned via a neural encoder, providing a more structured feature space for dimensionality reduction. Experiments on the MNIST, Fashion-MNIST, CIFAR-10, and STL-10 datasets demonstrate that the proposed method improves clustering performance and yields more interpretable embeddings than standard and extended t-SNE approaches.
Download

Paper Nr: 152
Title:

Detection of Anxiety and Depression from Social Media Text Using Natural Language Processing

Authors:

Muhammad Azhar, Adeen Amjad, Bilal Hussain, Chan Kit Ling, Taimoor Niaz and Muhammad Hamza Akbar

Abstract: The increasing global burden of anxiety and depression necessitates innovative computational approaches for early, non-intrusive detection. Social media platforms offer naturalistic data where individuals voluntarily express psychological states through text. This paper presents a comprehensive Natural Language Processing (NLP) pipeline for detecting linguistic correlates of anxiety and depression from social media text, a task distinct from clinical diagnosis. We systematically compare traditional machine learning models (Logistic Regression, SVM, Random Forest) and deep learning architectures (BiLSTM with Attention, fine-tuned BERT) across three benchmark datasets: SMHD, Dreaddit, and CLPsych 2015. Fine-tuned BERT achieves the highest macro-averaged F1 scores of 0.918, 0.909, and 0.862 respectively, outperforming traditional models by 7.1–17.3%. SHAP analysis identifies negative emotion, past-focused language, and somatic markers as the most predictive features. Cross-dataset validation reveals 15–25% performance degradation, highlighting the need for domain adaptation. Ethical considerations including false-positive risks and the gap between linguistic detection and clinical validity are discussed.
Download

Paper Nr: 168
Title:

Temporal Stability in Student Dropout Prediction Using Weekly Learning Analytics

Authors:

Tiago Franco, Paulo Alves, José Rufino, Maria de Fátima Pacheco and Nuno Ribeiro

Abstract: This study evaluates the temporal stability of student dropout prediction using Random Forest models optimized via Genetic Algorithms on weekly academic data. Based on real data from a Portuguese higher education institution, between 2017 and 2023, three models were optimized at distinct academic weeks and tested on 2023 students. The results indicate that hyperparameter configurations vary across snapshots, affecting performance. A cumulative dropout probability approach was used to assess students’ risk over time, showing a 10–15% improvement in early-warning precision. The findings support the implementation of adaptive, time-sensitive predictive models to enhance institutional interventions and student retention strategies.
Download

Paper Nr: 180
Title:

CRAFT: Clustered Regression for Adaptive Filtering of Training Data

Authors:

Parthasarathi Panda, Asheswari Swain and Subhrakanta Panda

Abstract: Selecting a small, high-quality subset from a large corpus for fine-tuning is increasingly important as corpora grow to tens of millions of datapoints, making full fine-tuning expensive and often unnecessary. We propose CRAFT (Clustered Regression for Adaptive Filtering of Training data), a vectorization-agnostic selection method for training sequence-to-sequence models. CRAFT decomposes the joint source-target distribution and performs a two-stage selection: (i) match the validation source distribution through proportional budget allocation across k-means clusters, and (ii) within each source cluster, select training pairs whose target embeddings minimize a conditional expected distance to the validation target distribution. Stage 1 is a stratified-sampling estimator of the source marginal, and Stage 2 is a regression toward the conditional mean of validation targets. We evaluate CRAFT on English-Hindi translation by selecting from 33 million NLLB sentence pairs and fine-tuning mBART-50 via LoRA, using an in-domain NLLB-derived validation set as both the selection target and the evaluation set. CRAFT achieves 43.34 BLEU, outperforming TSDS (41.21) by 2.13 points on the same candidate pool and encoder while completing selection over 40× faster. With TF-IDF vectorization, the entire pipeline completes in under one minute on CPU. TAROT achieves 45.61 BLEU, but CRAFT completes selection in 26.86 seconds versus TAROT’s 75.6 seconds, a 2.8× speedup.
Download

Paper Nr: 188
Title:

Data-Driven Multi-Horizon Taxi Demand Forecasting Using Transformer-Based Temporal Modeling

Authors:

Magesh Ram Rajakumar, Christine Markarian and Shadi Atalla

Abstract: Accurate short-term taxi demand forecasting is essential for efficient urban transportation, as it directly influences fleet allocation, passenger waiting times, and quality of service. This work studies multi-horizon taxi demand prediction as the task of learning a mapping from past observations to multiple future demand values under temporal dependence and uncertainty. A data-driven framework is developed using data from the New York City Taxi and Limousine Commission, where raw trip records are transformed into structured temporal representations through lag-based features, cyclical encoding, and statistical summaries. The framework evaluates sequence modeling architectures for multi-step prediction, such as recurrent models, in particular Long Short-Term Memory (LSTM) networks, and attention-based models, i.e., Transformer architectures. Evaluation across multiple forecasting horizons reveals that attention-based models consistently achieve lower prediction error, indicating their capability to capture the long-term temporal dependencies needed to predict multiple horizons. Conformal prediction and Monte Carlo Dropout are used to produce calibrated prediction intervals and point forecasts that represent demand variability under uncertainty. The findings indicate that successful multi-horizon forecasting requires global temporal interaction and explicit quantification of uncertainty. The proposed framework provides a systematic and scalable approach for real-world deployment in intelligent transportation systems.
Download

Paper Nr: 192
Title:

Exceptional Itemsets with Statistical Background Discovered by Evolutionary Computation

Authors:

Kei Arai and Kaoru Shimada

Abstract: This study proposes a framework for discovering, evaluating, and ranking Itemsets with Statistically Distinctive Backgrounds (ItemSBs) by combining GNMiner, an outcome-accumulation evolutionary method, with an Exceptional Model Mining (EMM)-inspired evaluation approach. Conventional EMM methods often rely on beam search, which may discard potentially promising candidates during level-wise pruning. To address this limitation, GNMiner was employed as an evolutionary rule-search method that accumulates diverse candidate ItemSBs during exploration. For post-hoc evaluation, we introduce an exceptionality measure, pdi f f , based on Fisher’s ztransformation, which quantifies differences in the correlation structure between a subset defined by ItemSB and its complement. Experiments conducted on housing-price and possum datasets demonstrate that the proposed framework can identify ItemSBs exhibiting strong deviations from the overall correlation structure and rank them using a unified scale. The results further suggest that pdi f f is useful for detecting localized relationships, particularly when global correlations are weak, while visual inspection remains important for interpreting practical differences.
Download

Paper Nr: 197
Title:

Corporate Reputation Exploration with Knowledge Graphs on Bipolar Extremes for Massive Online Consumer Review Datasets

Authors:

R. E. Loke and T. Schouten

Abstract: Marketeers and consumers are just two possible stakeholders having interest in finding out relevant corporate reputation patterns in huge datasets of online product reviews. In this paper, we strengthen knowledge graph application in such huge datasets by proposing a novel focused visualization strategy that explores data at the bipolar extremes of relevant aspects in corporate reputation dimensions. The visualization pipeline is operated with few-shot classifiers based on sentence transformers, REBEL (Cabot, 2021) and Neo4j. Results demonstrate efficient interactive visualization and exploration of generated graphs.
Download

Paper Nr: 224
Title:

Toward Reliable Automated Radiography: Improving Chest X-Ray Diagnosis through Multimodal Fusion and Label Cleaning

Authors:

Barbod Yadali Jamalouei, Pourya Jafari and Rayan Abri

Abstract: Automated chest X-ray diagnosis faces two fundamental challenges: most deep learning models use only visual features, ignoring the clinically rich text of radiology reports, and large-scale datasets like MIMIC-CXR suffer from systematic label noise introduced by automated annotation tools. Here, we propose a multimodal framework that addresses both challenges at the same time, by combining EfficientNet-B3 image features and Bio_ClinicalBERT text embeddings from the radiology reports, and a confidence-threshold label noise correction pipeline. We perform a systematic three-way ablation study on MIMIC-CXR for the five disease categories: Pleural Effusion, Atelectasis, Cardiomegaly, Pneumonia, and Pneumothorax. Our results indicate that: (1) incorporating text modality increases Macro AUC from 0.832 to 0.981 (+17.9%); (2) label cleaning using only image improves Precision by 3.97% but has a negligible effect on AUC; and (3) the combination of multimodal learning and label cleaning results in the best Precision of 0.893 with per-disease AUC ranging from 0.958 to 0.992. These findings establish that clinical report text is an underutilized resource in chest X-ray AI, and that data quality correction is most effective when paired with sufficiently expressive multimodal models.
Download

Paper Nr: 233
Title:

Agentic GraphRAG and Deterministic Schema Reconciliation for High-Compliance Domains: An LLMOps and FinOps Approach

Authors:

Marcelo Massashi Simonae, André Roberto Ortoncelli and Marlon Marcon

Abstract: The dispersal and interdependence of regulations pose significant challenges for Artificial Intelligence (AI) systems in high-compliance domains, where regulatory adherence requires absolute lexical precision and multi-hop relational reasoning. Traditional vector-based Retrieval-Augmented Generation (RAG) often fails in these scenarios, leading to structural hallucinations and a lack of institutional grounding. This paper proposes a scalable Agentic GraphRAG architecture structured under a comprehensive LLMOps Feature-Extraction-Inference (FTI) lifecycle. The core innovation is a deterministic Hybrid Vector-to-Graph coupling (Chunk-to-Entity Mapping) orchestrated by autonomous agents via the Model Context Protocol (MCP), ensuring strict referential integrity through batch schema synchronization. Experimental results on a normative academic benchmark demonstrate that while standard Vector RAG may offer marginal semantic fluency, the proposed Agentic GraphRAG definitively prioritizes auditability. Specifically, the architecture achieves a 70% improvement in Citation Rate compared to standard Vector RAG and establishes a robust Safe Abstention rate of 0.1800, effectively mitigating the risk of ungrounded generation. Furthermore, a FinOps evaluation across multiple Large Language Models (LLMs) validates the system’s cost-effectiveness, successfully optimizing the compliance-to-cost ratio for sustainable enterprise deployment.
Download

Paper Nr: 240
Title:

A Comprehensive Evaluation of Pre-Trained Language Models for Irony Detection in Tweets

Authors:

Mayara C. Marinho and Vinicius R. P. Borges

Abstract: Irony is a form of expressing the opposite meaning of the literal interpretation of a text. On social networks, users often employ irony in comments and discussions, posing challenges for various sentiment analysis tasks. This paper presents a comparative evaluation of different language models for detecting irony in tweets, covering multiple architectures: ULMFiT, pre-trained transformer encoders (BERT-base, XLM-RoBERTa, and ModernBERT), open-source LLMs, including Llama-3.1-8B, Gemma-7B, Qwen3-8B, and DeepSeek-LLM-7B. All models are assessed using a holdout strategy and F1-score for the ironic class on four publicly available corpora, including SemEval-2018 Tasks A and B. In addition, a Stratified 5-fold cross-validation was conducted to enable a statistical analysis to verify whether there is a significant performance difference between the models for each corpus. Results indicate that fine-tuned and quantized LLMs, particularly Gemma-7B, consistently achieved the highest F1-scores across most of the evaluated corpora.
Download

Area 6 - Databases & Data Management

Full Papers
Paper Nr: 196
Title:

Scalable Data Management for Smart Campus Digital Twin Applications

Authors:

Luís Carlos Casanova Afonso, Jo ao Rafael Almeida and José Luís Oliveira

Abstract: Understanding how students and staff move within university campuses is critical for optimizing space utilization, network resource allocation, and operational efficiency. Traditional approaches, such as manual surveys and static occupancy logs, fail to capture the complexity of campus mobility, while real-time monitoring introduces massive volumes of positioning data that are challenging to store, query, and analyze. Smart Campus Digital Twins promise to overcome these limitations by integrating Wi-Fi positioning, IoT sensing, and analytics, yet selecting an appropriate database management system remains an open problem: existing benchmarks target DevOps monitoring workloads and lack the mix of time-series and user-centric queries required for mobility analysis. In this work, we propose a workload taxonomy organizing Smart Campus DT queries by temporal service class, covering continuous, periodic, on-demand, and event-triggered query patterns, and release a reproducible benchmark built on a synthetic dataset of 58 million Wi-Fi records and 20 queries spanning the taxonomy. We evaluate six DBMSs covering relational, time-series, and columnar architectures under default and tuned configurations as well as concurrent ingestion with multi-client dashboard reads. Results show that ClickHouse excels across the full taxonomy, QuestDB enables ultra-low-latency visualization, TimeScaleDB offers SQL-compatible improvements, and a purpose-built time-series engine fails on most user-centric queries.
Download

Paper Nr: 214
Title:

Read Fast, Write Carefully: Empirical Evidence on the Performance Trade-Offs of Denormalization in PostgreSQL under TPC-H Workloads

Authors:

Luís Mendonça, Pedro Martins, Filipe Madeira and Maryam Abbasi

Abstract: Relational database normalization is a cornerstone of sound schema design, yet practitioners often denormalize schemas to improve read performance. This paper presents a controlled empirical study of PostgreSQL 15 under a row-store configuration, covering six TPC-H queries (Q1, Q3, Q6, Q10, Q12, Q15), five scale factors (SF 0.1 to SF 5), and three schema variants—normalized (3NF), denormalized (DENORM), and indexed normalized (IDX). The study includes 1,350 executions, with 25 repeated runs per configuration, instrumented with per-run buffer I/O using EXPLAIN (ANALYZE, BUFFERS). Under the tested configuration, aggressive full denormalization degraded median read performance by 1.95–2.91× at small scale factors, primarily because wider rows reduced page density and increased block reads by 6.8–8.3×. A secondary finding is a stability paradox: denormalized warm runs exhibited lower coefficient of variation (4.8–5.4%) than some normalized counterparts. This did not indicate greater efficiency; rather, the wider table permanently exceeded the shared buffer pool, making executions consistently disk-bound. Per-run block counts remained constant across warm runs, and the near-zero correlation (r = −0.065) between the hit/read split and execution time indicates that total I/O volume, rather than cache state, dominated execution cost. Composite indexes on the normalized schema matched or outperformed the unindexed 3NF baseline in the tested scenarios and reduced variance by up to 82% at SF 0.5. Scaling analysis further showed that Q12 grew super-linearly (9.25× from SF 1 to SF 5, against the expected 5×), consistent with hash-join work-file spill under the default 4 MB work mem. These findings suggest that, for this PostgreSQL row-store setup, composite indexing should be evaluated before aggressive schema flattening, and that low variance in a flat-table scan may indicate chronic disk-boundedness rather than efficient caching.

Paper Nr: 237
Title:

Racial Inequality in Brazilian Computing Education: An Analysis of Black Students Using Census Data

Authors:

Francisco Carlos Monteiro Souza, Marlon Marcon, André Roberto Ortoncelli, Rodolfo Adamshuk Silva and Alinne Cristinne Corrêa Souza

Abstract: The expansion of higher education in Brazil has unfolded alongside persistent racial and social inequalities. Affirmative action policies improved access for marginalized groups, but computing programs remain marked by pronounced disparities. Despite the implementation of racial quotas and financial aid, little is known about how these policies affect the academic trajectories of Black students in computing. Few studies explore how region, institution type, and quota usage shape access, persistence, and outcomes. This study analyzes Higher Education Census data from 2010–2024 to examine entrants, enrollment, and graduation patterns of Black students in computing, assessing the impact of affirmative action and regional/institutional disparities. A descriptive and quantitative approach was adopted, following CRISP-DM for data mining. Census microdata were processed to identify participation patterns disaggregated by region, institution type, and modality. Findings confirm that affirmative action expanded access for Black students. However, significant gaps persist in graduation rates and regional representation. The study provides empirical evidence on racial inequality in computing education, offering insights to guide both academic debate and public policies. It contributes to the Information Systems field by examining how large-scale educational data can inform equity-oriented decision-making, with potential implications for the design of systems that support student retention and inclusion.
Download

Short Papers
Paper Nr: 90
Title:

Modeling Higher-Order Relationships in the Context of Big Data: Methods, Applications, and Prospects

Authors:

Huijun Zheng, Qing Zhang, Longyun Wang, Min Wu, Wenchuan Yang and Xin Lu

Abstract: The advent of big data has significantly enriched network science, catalyzing the evolution of modeling and analytical techniques for higher-order relationships within complex systems. This progress unlocks novel pathways for deciphering the underlying dynamics of such systems and holds substantial promise for practical applications. Centering on the intricate connections among system components, this survey begins by elucidating the fundamental concepts, specifically differentiating between higher-order interactions and higher-order dependencies. It then offers a systematic review of prevailing modeling frameworks, including those leveraging simplicial complexes, hypergraphs, network motifs, and higher-order Markov processes. Subsequently, the paper presents a comprehensive overview of current applications, highlighting cutting-edge developments in social, biological, and engineering-physical systems. The study concludes by critically assessing the prevailing challenges from both theoretical and applied standpoints, and finally, outlines potential directions for future research in this rapidly evolving field.
Download

Paper Nr: 92
Title:

Research on Core Technology Identification Methods Based on High-Order Dependency Metrics

Authors:

Qing Zhang, Siyu Lai, Huijun Zheng, Longyun Wang, Wenchuan Yang, Min Wu and Xin Lu

Abstract: Core technology identification is critical for securing technological high ground and optimizing the allocation of innovation resources, and complex network-based analytical approaches have become an important research avenue. Introducing higher-order dependency metrics enables a comprehensive characterization of the multi-node interactions and nonlinear dependencies among technology nodes, thereby enhancing the accuracy and interpretability of core technology identification. Based on complex network theory, different-order Markov dependencies are incorporated into the technology citation network to construct a higher-order dependency network. A metric system is then established across three dimensions-local structure, global network features, and diffusion dynamics-to evaluate node centrality, and its performance is compared with conventional first-order metrics to assess the advantages of higher-order dependencies in identifying core technologies. The results indicate that higher-order dependency metrics can effectively identify technologies occupying core positions in the citation network, providing a more interpretable and reliable framework for core technology identification.
Download

Paper Nr: 154
Title:

Finding the Sweet Spot: Query Cost and Load Distribution for Spatio-Temporal Queries in Peer-to-Peer Systems

Authors:

Zhan Ye, Laurent Yeh and Iulian Sandu Popa

Abstract: Spatio-Temporal queries over distributed hash tables (DHTs) remain difficult because they pull the design in two opposite directions: preserving key continuity enables efficient range scans, while strong load balancing tends to disperse responsibility and fragment contiguous ranges. We consider spatio-temporal queries in a peer-to-peer overlay (i.e., a Chord-like ring). At one design end, each record is mapped to a totally ordered 1D overlay key via a 3D space-filling-curve encoding and placed Order-preservingly. This setting yields low query routing cost and naturally supports successor-chain node scans. Yet it exposes severe hotspot skew. Conversely, global dispersal mechanisms, such as virtual nodes in Dynamo-style token rings, improve balance but can drastically increase the number of nodes touched by an overlay key-interval scan. We propose SmartNode, a lightweight local compensation mechanism. SmartNode performs minimal boundary splits only at overloaded key-run boundaries and spills the tail interval to the successor, flattening hotspots while confining fragmentation to boundaries rather than dispersing the entire key space. Using the Geolife trajectory data set, we empirically show that this continuity-first design point achieves near-uniform load while keeping the cost of overlay key-interval scans close to the Order-preserving baseline across multiple network scales and query densities. These results position local boundary compensation as a practical alternative to global dispersal for spatio-temporal query workloads on DHT overlays.
Download

Paper Nr: 186
Title:

MLBRS: A Multi-Layer Behavioural Risk Scoring Framework for Insider Threat Detection

Authors:

V. L. Kartheek, Aayush Shah, Rishav Jain, R. Gururaj and Subhrakanta Panda

Abstract: Insider threats remain difficult to detect because malicious actions often resemble legitimate user behaviour and may evolve gradually over time. This paper presents MLBRS, a multi-layer behavioural risk scoring framework that combines rule-based scoring, statistical deviation analysis, and Isolation Forest-based anomaly detection to generate continuous employee-level risk scores. The framework integrates behavioural indicators, personalised deviation modelling, and multivariate anomaly detection to identify both abrupt and gradual behavioural changes. Due to the limited availability of publicly accessible datasets containing database-query-level insider threat activity, a synthetic dataset was constructed to simulate organisational behaviour with temporal consistency, multiple employee roles, and diverse attack scenarios. Existing insider-threat datasets primarily capture system-level activity and do not adequately represent database interactions. Experimental evaluation demonstrates consistent detection performance, achieving an ROC-AUC of 0.978 and an F1-score of 0.88 on the synthetic dataset. Additional cross-dataset evaluation using CERT-derived behavioural traces shows reduced but stable performance under less aligned behavioural conditions. The results indicate that MLBRS provides an interpretable and scalable approach for behavioural insider threat detection across heterogeneous activity patterns.
Download