Continual Learning Approach With Adaptive Memory Mechanism for Predictive Process Monitoring in Dynamic Business Environments

Predictive process monitoring (PPM) faces challenges when process patterns evolve over time, as prediction models experience performance degradation due to concept drift while losing accuracy on previously learned process variants through catastrophic forgetting. This study addresses these challenges in outcome-oriented PPM by integrating continual learning principles with process mining domain knowledge. We propose an adaptive memory mechanism that extracts process structure information through activity relationship analysis, categorizing relationships between activities as “Always Follows,” “Sometimes Follows,” or “Never Follows.” The mechanism computes relation entropy vectors to characterize process patterns and uses cosine similarity between current and historical process patterns to adaptively weight memory buffers during model training. This allows the model to weight historically relevant process knowledge based on pattern similarity. We evaluate the mechanism using stability, plasticity, and trade-off metrics on the Business Process Intelligence Challenge 2015 dataset across multiple temporal splits. Results show that the adaptive memory mechanism achieves stability of 0.87 ± 0.06, plasticity of 0.84 ± 0.11, and trade-off of 0.85 ± 0.08, achieving competitive performance compared to established baseline methods. The mechanism provides explainability through visualizations that demonstrate how the model weights different memory buffers during training. Validation using concept drift detection techniques confirms that the mechanism’s behavior aligns with observable process changes, indicating adaptation to evolving process patterns while reducing catastrophic forgetting.