On-the-fly Event Disambiguation via Alignments

This dataset contains the supplmentary material to the article"On-the-fly Event Disambiguation via Alignments".

It includes:

Two ambiguous event logs in the XES format extended with data uncertainty (files BD_ambiguous.xes and SH_ambiguous.xes), used to evaluate the disambiguation algorithm proposed in the paper. Each ambiguous event log is complemented by a gold-standard log (files BD_gold_standard.xes and SH_gold_standard.xes), which corresponds to its unambiguous version. The traces in each log describe 15,000 executions of a process:

Traces in the BD log describe executions of a blood donation process, synthesized from 32 real executions of the process in an IoT-augmented setup.

Traces in the SH log describe executions of a smart home process, derived from the BP-Meets-IoT 2020 challenge dataset (see details here), by first filtering for activities related to kitchen and eating routines, and then using the inductive miner to derive a process model.

Sets of ambiguous events in the logs are: [{Check blood drawing machine, Take out samples}, {Disinfect injection site, Apply tourniquet, Insert needle, Remove tourniquet, Remove needle}] and [{Go fridge, Go dining table}, {Get cold warm food, Go kitchen shelf, Get food from fridge}, {Go micro, Get bread}].

The original traces are stochastically perturbed to reflect real IoT-augmented monitoring settings by randomly skipping process events with a probability of 20%. This was done using Java's java.util.Random without fixing a seed. Hence, perturbation is stochastic and not tied to a specific deterministic replication. Each released log corresponds to one concrete instantiation of this stochastic generation process. This results in traces being potentially non-conforming to the process specification.

Moreover, unpredictability of hand hygiene events in healthcare scenarios is reflected by introducing additional Perform hand hygiene events at random points, following the same strategy as for the skipping of process events. This results in traces being potentially non-conforming to the process specification.

A Petri net describing the normative process model for each log. This is encoded in a respective file (BD_model.pnml and SH_model.pnml).

The results of the algorithm evaluation, which aimed to assess the algorithm's ability to correctly disambiguate events with latencies acceptable for online monitoring settings. These are presented in files BD_results.xslx and SH_results.xlsx.

The files report on the runtimes of the algorithm on both ambiguous and gold-standard traces, and on the disambiguation accuracy and fitness of the disambiguated traces.

Runtimes are presented both cumulatively and non-cumulatively.

Disambiguated logs resulting from a baseline null model disambiguation approach applied to both BD_ambiguous.xes and SH_ambiguous.xes (files Baseline_BD_disambiguated.xes and Baseline_SH_disambiguated.xes). The null model disambiguation approach operates without access to the process model or any contextual information. Consequently, it cannot assess the plausibility of competing interpretations and treats them as equally plausible. Therefore, it disambiguates events by randomly selecting one among the possible interpretations. For reproducibility, the random choice is performed using a fixed seed (42).

Disambiguation accuracy and fitness drop relative to the baseline approach are included in file Baseline_results.xlsx.

The source code of the algorithm implementation for the on-the-fly event disambiguation via alignments is available here (GitHub repository).