A centroid aware swarm intelligence approach for behavior oriented trace clustering
Аннотация
Complex event logs contain diverse execution patterns that can make discovered process models difficult to interpret. This study introduces PSOCluster, a behavior-oriented trace-clustering framework that integrates an interpretable binary behavioral representation with centroid-aware particle swarm optimization. Each trace is encoded using selected Declare-style rules for activity occurrence, ordering, immediate succession, and co-occurrence. Support filtering, entropy-based ranking, and Jaccard pruning retain a compact set of informative features. Each particle represents a complete configuration of K centroids, jointly optimized as points in the behavioral feature space. The search retains the best complete configuration for each particle and maintains centroid-indexed social memories based on cluster-level quality. Decreasing inertia, Gaussian mutation, and centroid repulsion support exploration and discourage centroid overlap. Candidate configurations are evaluated using an adaptive Silhouette–Density objective that combines normalized neighborhood density with silhouette and progressively increases the emphasis on silhouette during the search. Replay fitness and precision are measured on the process models discovered after clustering. On a laboratory event log of 18,891 traces, feature selection reduced 2128 candidate rules to 67 features (96.85%). Across 30 random seeds, PSOCluster achieved a mean size-weighted process-model F1-score of 0.98 (SD = 0.04). In the laboratory comparison, it produced six clusters with full trace coverage and a higher F1-score than the evaluated full-coverage methods. On four public event logs, it attained the highest reported weighted F1-score among the evaluated methods on each log, with an unweighted mean of 0.9537.