Spatial-temporal knowledge graph network for event prediction

Abstract

Predicting multiple concurrent events has a remarkable effect on understanding social dynamics and acting in advance to reduce damage. (1) From the perspective of spatial connection, trans-regional implication, which means the cause of the incident is not local but somewhere else, is an important reason for the occurrence of events. However, existing works neglect to model this spatial influence and only leverage the local information for event prediction. (2) From the perspective of temporal connection, future events are triggered by the continuous evolution of the region. Nonetheless, most studies assign events to different timestamps and recognize their sequential patterns, ignoring the continuity of the evolution process. To tackle the above two problems, we propose a spatial and temporal knowledge graph neural network (STKGN). Specifically, to construct the cross-regional connection, we propose a novel spatial-temporal event graph, where each region is denoted as a node and trans-regional influences are reflected by bidirectional edges. To simulate the continuously evolving process, we propose an event-driven memory network to represent the state of each entity and continually update the state embeddings by emerging events. Then we use a broadcast network to spread the local changes in the graph to obtain high-order reasons like the trans-regional implication. Extensive experiments on two real-world datasets demonstrate that STKGN achieves significant improvements over state-of-the-art methods. Further analysis shows the interpretability of the trans-regional implication.