| 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. |