Efficient Directed Hypergraph Network for Unsupervised Traffic Anomaly Detection

Anomaly detection in traffic flow is crucial for ensuring the safety and stability of intelligent transportation systems. However, many transformer-based deep learning models suffer from extremely high computational complexity, making them unsuitable for real-time anomaly detection in resource-constrained terminal devices. Therefore, this article presents an efficient and fully unsupervised framework for traffic anomaly detection, termed the directed hypergraph message passing network (DHMPN). By integrating a directed hypergraph-based message passing module with bidirectional gated linear units (Bi-GLUs), DHMPN effectively captures complex spatiotemporal dependencies while maintaining low computational overhead, thereby addressing the practical demands of real-time anomaly detection. To enable accurate anomaly identification, the learned traffic representations are further modeled using a conditional normalizing flow, which facilitates precise density estimation and probabilistic anomaly scoring. Extensive experiments on real-world traffic benchmark datasets demonstrate that DHMPN significantly outperforms both traditional machine learning approaches and state-of-the-art graph-based deep learning models in anomaly detection tasks. Moreover, DHMPN achieves competitive performance in traffic forecasting, further validating the effectiveness of the proposed spatiotemporal encoding architecture.