Probabilistic Metaheuristic Ensembles for Scalable Process Mining and Event Log Prediction in Enterprise Systems

• Huy Tran

  Thai Nguyen University of Technology, Dang Tat Road 56, Department of Information Systems Engineering, Thai Nguyen City, Thai Nguyen Province, Vietnam

  Author

• Phuc Le

  Can Tho University, Khu 2, 3/2 Street 128, Department of Software Technology and Analytics, Can Tho, Can Tho Municipality, Vietnam

  Author

Enterprise information systems routinely record fine-grained event logs that capture the execution of business processes across heterogeneous applications, organizational units, and time horizons. Process mining and event log prediction aim to extract behavioral models and predictive signals from such logs, yet practical deployments face scale, noise, incompleteness, nonstationarity, and the need to balance accuracy with interpretability and operational constraints. This paper develops a probabilistic metaheuristic ensemble perspective for scalable process discovery, conformance-oriented optimization, and prefix-based event log prediction. The central premise is that no single metaheuristic consistently dominates across logs, process variants, and objective trade-offs, so ensembles should allocate computation adaptively under uncertainty. We formalize process mining tasks as stochastic multiobjective optimization problems over candidate model spaces and show how probabilistic portfolios can coordinate diverse search operators while providing calibrated uncertainty estimates on both model quality and predictions. The proposed framework combines Bayesian performance modeling, bandit-style resource allocation, and distributional search updates to guide discovery and prediction under bounded latency budgets. Scalability is addressed through approximate fitness estimation, incremental evaluation on streaming traces, and distributed asynchronous execution that preserves probabilistic accounting of solver contributions. For prediction, the paper connects discovered process structure with probabilistic sequence and time models via mixture formulations that support uncertainty-aware next-event and remaining-time inference. The discussion emphasizes enterprise requirements, including drift handling, governance, and evaluation practices that separate offline quality from online utility.

• PDF