A Hybrid Framework for LLM-Assisted Concept Drift Analysis in Process Mining

Business processes evolve continuously over time. Variations in workload, organizational practices, seasonality, or external conditions may alter the way a process is executed, making concept drift (CD) an important issue in process mining (PM). Existing drift detection methods locate changes but say little about what happened and offer no natural-language interaction, which limits their use by non-experts. Large Language Models (LLMs) show strong capabilities in natural language, but their output may lack support from the event data. We propose a hybrid framework for drift detection in PM by combining PM techniques and LLM capabilities. The event log is first divided into temporal non-overlapping windows of configurable size (40 for the synthetic log, 80 for the real-life log). Sublogs are abstracted and then compared by an LLM to extract candidate changes, which are validated by Kolmogorov–Smirnov and Mann–Whitney tests (α=0.05). Validated changes are stored in a Change Log and organized into a classified Change Tree, which is used as the knowledge source of a Retrieval-Augmented Generation (RAG) component for natural-language exploration. On a synthetic log, the injected Risk Assessment insertion was validated (p<0.001), and ProDrift localized a drift at the same position. On the real-life SEPSIS log, three changes were reported and compared with ProDrift results.