New Hypergraph Network Boosts Next Activity Prediction in Complex Processes

Predicting the next activity in complex service-oriented processes is crucial for preventing delays and managing risks. Current methods often struggle with object-centric event logs (OCELs), which represent events shared by multiple business objects, leading to a loss of cross-object context or inefficient encoding of multi-object interactions. This paper introduces EHHN, an Event-driven Heterogeneous Hypergraph Network, designed to overcome these limitations. EHHN models each prediction prefix as a heterogeneous hypergraph, effectively binding co-participating objects and grouping primary object lifecycle events. Its dual-stream architecture processes both micro-spatial event-driven object state evolution and macro-evolution temporal dynamics using global prototypes. This integrated approach allows EHHN to achieve superior accuracy and F1-scores on four public OCEL benchmarks, outperforming nine baselines by up to 8.1 and 12.4 percentage points. Furthermore, EHHN significantly reduces peak GPU memory usage, demonstrating its efficiency. The code is publicly available.