Uncertainty-aware traffic accident risk prediction via multi-view hypergraph contrastive learning

Abstract

Traffic accident prediction is crucial for maintaining safety in smart cities. Accurate prediction can significantly reduce casualties and economic losses, while alleviating public concerns about urban safety. However, achieving this is challenging. First, accident data exhibits twofold imbalances: (i) a class imbalance between accident occurrence and non-occurrence, and (ii) a spatial distribution imbalance among different regions. Second, sporadic traffic accidents result in sparse supervised signals, limiting the spatial–temporal representations of conventional deep models. Lastly, the Gaussian assumption underlying the previous deterministic deep learning models is unsuitable for accident risk data characterized by dispersed and many zeros. To address these challenges, we propose an Uncertainty-aware spatial–temporal multi-view hypergraph contrastive learning framework for Traffic accident risk prediction (TarU). This framework not only jointly captures local geographical spatial–temporal and global semantic dependencies from different views, but also parameterizes the probabilistic distribution of accident risk to quantify uncertainty. Particularly, a hypergraph-enhanced network and an auxiliary contrastive learning architecture are designed to enhance self-discrimination among regions. Extensive experiments on two real-world datasets demonstrate the effectiveness of TarU. The proposed framework may also be a paradigm for addressing spatial–temporal data mining tasks with sparse labels.