New Framework for Continual Fine-Tuning of Foundation Models in Process Monitoring

Predictive Process Monitoring (PPM) models often struggle in real-world scenarios where underlying processes change over time, a phenomenon known as concept drift. While online continual learning methods exist, they typically train new models from scratch, leading to a "cold-start" problem. Foundation Models (FMs) offer a promising solution due to their pre-trained knowledge, but their application in continual fine-tuning for process mining has been largely unexplored. A new framework, COMPASS (Continual Online foundation Model-based PPM with Adaptive SubSpaces), has been developed to tackle this. COMPASS is designed for online continual fine-tuning of FMs in PPM. It incorporates an adaptive loss-plateau drift detection mechanism to automatically identify when process changes occur in event streams. The framework also maintains a unified knowledge subspace, integrating both pre-trained FM knowledge and task-specific adaptations. Evaluations across various synthetic and real-world concept drift scenarios, using different model backbones, show that COMPASS significantly outperforms three state-of-the-art non-FM competitors and two baseline update strategies. Its performance is particularly strong in cases with recurrent drift and complex, long-running processes, all while maintaining acceptable computational overhead.