Adaptive Event Disambiguation via Alignments with Deferred Decisions

This dataset contains the supplementary material to the article “Adaptive Event Disambiguation via Alignments with Deferred Decisions”, submitted to the journal Process Science. datasets/: the two synthetically generated datasets used in the evaluation, each in its own folder (blood_donation, smart_home). Each folder contains: runs.txt: the normative process behavior as a set of reference runs (one run per line), i.e. the model log against which traces are aligned. Activities are encoded as single symbols. labels.json: the mapping between activity labels and the symbols used in runs.txt. ambiguous.xes: the ambiguous event log, i.e. a set of traces in which events affected by ambiguity carry all their competing candidate activity labels. gold_standard.xes: the corresponding unambiguous log, used as ground truth for measuring disambiguation accuracy. Each log describes 15,000 executions of a process: Blood donation (BD): a blood-donation process synthesized from 32 real executions in an IoT-augmented lab setup. Smart home (SH): a smart-home process derived from the BP-Meets-IoT 2020 challenge dataset (details) by filtering for kitchen/eating-routine activities and discovering a model with the inductive miner. The sets of mutually ambiguous activities are {Check blood drawing machine, Take out samples} and {Disinfect injection site, Apply tourniquet, Insert needle, Remove tourniquet, Remove needle} for BD, and {Go fridge, Go dining table}, {Get cold warm food, Go kitchen shelf, Get food from fridge}, {Go micro, Get bread} for SH. The generated traces are stochastically perturbed to reflect IoT-based monitoring imperfections: process events are randomly skipped (simulating activities undetected by the IoT setup), and, for the healthcare scenario, additional hand hygiene events are inserted at random points (reflecting their unpredictable occurrence). Perturbation uses Java’s java.util.Random without a fixed seed, so each released log is one concrete instantiation of this stochastic process; the resulting traces may be non-conforming to the normative behavior. The third dataset used in the paper, HR recruitment (HR), is a publicly available uncertain event log and is not redistributed here; it is available at its original repository (description: HaifaUniversityBPM/traffic-data-to-event-log). Only its evaluation results are included (see below). results/: the evaluation results for all three datasets (X ∈ {BD, SH, HR}), one workbook per algorithm configuration: X_adaptive.xlsx: the adaptive algorithm (ABD in the paper). X_maxk_y.xlsx: the fixed-bound algorithm (FBD) with exploration bound maxK = y, for y = 1..5. X_baseline.xlsx: the random-choice null-model baseline (RND), which has no access to the reference behavior and selects uniformly among the candidate labels of each ambiguous event. The random choice uses a fixed seed for reproducibility. Each workbook reports disambiguation quality and runtime for that configuration, organized into the following sheets: Summary: headline figures: dataset, algorithm (and maxK for FBD), number of events and traces, all-events disambiguation accuracy, mean fitness of the gold-standard and of the disambiguated traces, mean fitness drop, and mean runtime (per trace and per event). AllEvents: one row per event comparing the gold-standard label with the disambiguated label (Are equal); the basis of the all-events accuracy. Results: per-trace accuracy (total events, number of mismatches, accuracy). Fitness: per-case fitness of the gold-standard and disambiguated traces (run/trace length, synchronous and asynchronous moves, fitness) and the per-case fitness drop. Runtime_Ambiguous: per-trace runtime (ms) on the input log, with one column per measurement repeat and the average. The source code implementing the fixed-bound and adaptive algorithms is available in this GitHub repository.