GH-DSTPP: Global Context Awareness and Hawkes-Guided Attention for Spatio-temporal Diffusion Point Processes

The Spatio-temporal Diffusion Point Processes (DSTPP) model uses diffusion modeling to learn joint spatio-temporal event distributions. It overcomes the long-standing limitation of spatio-temporal independence assumptions in traditional point process models. However, its capacity to model dependencies governed by global temporal trends and underlying domain priors remains limited. Therefore, we propose a Global Context Awareness and Hawkes-Guided Attention framework for Spatio-temporal Diffusion Point Processes (GH-DSTPP). Specifically, we introduce a bidirectional LSTM coupled with multi-head self-attention to extract latent global temporal trends from historical events, thereby complementing DSTPP's local event-level modeling. Furthermore, to better capture the inherent physical dynamics and causal dependencies in point processes, we design a Hawkes-guided Attention Mechanism. It filters out irrelevant events to generate a more focused and meaningful attention distribution. Extensive experiments are conducted across five datasets: Earthquake, COVID-19, Citibike, HawkesGMM, Pinwheel. The results demonstrate GH-DSTPP's competitive or improved temporal likelihood and competitive, dataset-dependent spatial performance and its ability to capture scenario-adaptive spatio-temporal interactions. Source code is available at https://anonymous.4open.science/r/GH-DSTPP/ .