Quantitative Evaluation of Process Discovery Enhancement Using Clustering-Based Techniques
Year: 2026
Type: conference-paper
Abstract: Business process discovery aims to reconstruct accurate models of organizational workflowsfrom event logs, yet classical algorithms often struggle with noisy and heterogeneous data,producing overly complex or imprecise models. This paper presents a systematic, quantitative evaluationof clustering-based preprocessing as a means to enhance process discovery. Using the BPIChallenge 2014 incident management log, we applied comprehensive feature engineering to extractstructural, temporal, and behavioral attributes, followed by multiple clustering techniques (K-Means,DBSCAN, HDBSCAN, GMM). Within each cluster, four established discovery algorithms (Alpha Miner,Heuristic Miner, Inductive Miner, ILP Miner) were executed from the PM4Py Python libraryand evaluated using fitness, precision, generalization, and simplicity metrics. Results show that clusteringconsistently improved model simplicity and precision, while fitness remained stable and generalizationtended to decrease. The extent of improvement depended on both the clustering methodand the discovery algorithm: K-Means with higher cluster numbers enhanced fitness and simplicityfor Alpha Miner; HDBSCAN improved precision and simplicity for Heuristic Miner; InductiveMiner benefited from K-Means even with small cluster numbers; whereas ILP Miner models remainedhighly complex regardless of clustering. These findings highlight the trade-off between interpretabilityand generalization, and demonstrate that clustering-based preprocessing can yield moretransparent and actionable process models.
DOI: 10.4028/p-3geynn
Authors: Erika Baksáné Varga, Attila Baksa
Institution: University of Miskolc
Subfield: Management Information Systems
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7172329237