articles

Existing Process Mining techniques often exhibit low robustness to common real-world data issues, such as noisy traces and incomplete event logs. This paper presents the development and validation of an integrated three-stage approach that combines statistical anomaly filtering, probabilistic and temporal reconstruction of missing events, and adaptive process model synthesis. The study addresses the following tasks: a critical review of classical process discovery algorithms; formalization of filtering methods based on Isolation Forest and event reconstruction using probabilistic and temporal metrics; and the design of an adaptive mechanism for selecting the noise threshold based on the normalized entropy of subprocess variability. The approach is implemented as a Python software module using the pm4py, scikit-learn, and NumPy libraries. Experiments on synthetic datasets generated with varying noise levels and proportions of missing events confirm the robustness of the proposed method. The results are evaluated using the Fitness, Generalization, Simplicity, and F1-score metrics and compared against the Alpha Miner, Heuristics Miner, and Inductive Miner algorithms. The proposed approach yields a statistically significant improvement in the quality of the resulting process models under high noise and log incompleteness, providing a basis for robust business process analytics systems capable of operating on data from real information systems.