Abstract
Concept drift presents major challenges for Predictive Process Monitoring (PPM), as predictive models often struggle to make accurate predictions on traces that differ from the event log they were originally trained on. Nevertheless, some background knowledge can be mined from the drifting traces and be used for improving PPM algorithms. In this paper, we propose a general Neuro-Symbolic framework for multi-attribute suffix prediction that explicitly contextualizes the neural predictions with both declarative and procedural background knowledge. Our framework combines the ability to learn from historical data (the neural component) with the capacity to reason over background knowledge (the symbolic component), aiming to combine the robustness of neural networks with the compliance-checking capabilities of symbolic knowledge. This framework enables more flexible suffix prediction, better suited to the variability of real-world processes, which continuously evolve due to concept drift, and may exhibit new behavioral variants over time. The experimentation on real event logs with a natural concept drift show that incorporating symbolic knowledge, generally improves the performance of suffix prediction.