HDM-GNN: A Heterogeneous Dynamic Multi-view Graph Neural Network for Crime Prediction
Binbin Zhou, Hang Zhou, W. Wang, Liming Chen, Jianhua Ma, Zengwei Zheng
Abstract
Smart cities have drawn a lot of interest in recent years, which employ Internet of Things (IoT)-enabled sensors to gather data from various sources and help enhance the quality of residents’ life in multiple areas, e.g. public safety. Accurate crime prediction is significant for public safety promotion. However, the complicated spatial-temporal dependencies make the task challenging, due to two aspects: 1) spatial dependency of crime includes correlations with spatially adjacent regions and underlying correlations with distant regions, e.g. mobility connectivity and functional similarity; 2) there are near-repeat and long-range temporal correlations between crime occurrences across time. Most existing studies fall short in tackling with multi-view correlations, since they usually treat them equally without consideration of different weights for these correlations. In this paper, we propose a novel model for region-level crime prediction named as Heterogeneous Dynamic Multi-view Graph Neural Network (HDM-GNN). The model can represent the dynamic spatial-temporal dependencies of crime with heterogeneous urban data, and fuse various types of region-wise correlations from multiple views. Global spatial dependencies and long-range temporal dependencies can be derived by integrating the multiple GAT modules and Gated CNN modules. Extensive experiments are conducted to evaluate the effectiveness of our method using several real-world datasets. Results demonstrate that our method outperforms state-of-the-art baselines. All the code are available at https://github.com/ZJUDataIntelligence/HDM-GNN.