DyG-Pred: Spatial-Temporal Dynamic Graph Replay for Event Prediction With Gated Explicit-Implicit Graph Dependency Fusion

Abstract

Event prediction based on dynamic graph modeling captures the evolving relationships among events, entities, and environments, providing effective support for intelligent risk warning, resource allocation, and decision-making, particularly in public safety. Existing dynamic graph methods generally assume a fixed number of events (nodes) across time, whereas real-world event graphs continuously evolve with newly emerging events. This limitation hinders the modeling of event evolution and the accurate discovery of implicit event dependencies. To address this issue, we propose DyG-Pred, a dynamic event prediction framework based on Dynamic Link-to-Graph Replay Tracking with explicit–implicit gated graph fusion. First, an explicit event graph is constructed by extracting event link schemas from the observational event graph to capture static semantic relationships among events. Second, a spatial-temporal replay mechanism progressively constructs an implicit event graph by jointly modeling the evolution of nodes and edges as the graph size dynamically increases. Third, an explicit–implicit gated graph fusion module adaptively integrates the heterogeneous graph representations through dynamic weighting for final event prediction. Experimental results demonstrate that the proposed replay mechanism effectively captures long-range implicit event dependencies. By adaptively fusing these dependencies with the stable semantic information of the explicit event graph, DyG-Pred achieves excellent event prediction performance while maintaining strong generalization across large-scale datasets from different domains.