DriftXMiner: A Resilient Process Intelligence Approach for Safe and Transparent Detection of Incremental Concept Drift in Process Mining

Processes supported by process-aware information systems are subject to continuous and often subtle changes due to evolving operational, organizational, or regulatory factors. These changes, referred to as incremental concept drift, gradually alter the behavior or structure of processes, making their detection and localization a challenging task. Traditional process mining techniques frequently assume process stationarity and are limited in their ability to detect such drift, particularly from a control-flow perspective. The objective of this research is to develop an interpretable and robust framework capable of detecting and localizing incremental concept drift in event logs, with a specific emphasis on the structural evolution of control-flow semantics in processes. We propose DriftXMiner, a control-flow-aware hybrid framework that combines statistical, machine learning, and process model analysis techniques. The approach comprises three key components: (1) Cumulative Drift Scanner that tracks directional statistical deviations to detect early drift signals; (2) a Temporal Clustering and Drift-Aware Forest Ensemble (DAFE) to capture distributional and classification-level changes in process behavior; and (3) Petri net-based process model reconstruction, which enables the precise localization of structural drift using transition deviation metrics and replay fitness scores. Experimental validation on the BPI Challenge 2017 event log demonstrates that DriftXMiner effectively identifies and localizes gradual and incremental process drift over time. The framework achieves a detection accuracy of 92.5%, a localization precision of 90.3%, and an F1-score of 0.91, outperforming competitive baselines such as CUSUM + Histograms and ADWIN + Alpha Miner. Visual analyses further confirm that identified drift points align with transitions in control-flow models and behavioral cluster structures. DriftXMiner offers a novel and interpretable solution for incremental concept drift detection and localization in dynamic, process-aware systems. By integrating statistical signal accumulation, temporal behavior profiling, and structural process mining, the framework enables fine-grained drift explanation and supports adaptive process intelligence in evolving environments. Its modular architecture supports extension to streaming data and real-time monitoring contexts.