DPSN-STHA: A dynamic perception model of similar nodes with spatial-temporal heterogeneity attention for traffic flow forecasting

Abstract

Precisely capturing spatial-temporal feature correlations represents an effective approach for improving the traffic flow prediction performance. However, accurate capturing of spatial-temporal features in traffic systems faces certain challenges, such as long-range correlations and node heterogeneity. To overcome these issues, this paper introduces a novel traffic flow prediction model that incorporates spatial-temporal heterogeneous attention, allowing for dynamic perception of similar nodes. In the proposed model, a filter network dynamic parameter memory generator is used for real-time parameter adjustment, which assigns greater weights to nodes with higher similarity to mitigate spatial-temporal heterogeneity. In addition, a similarity-based node computation method, which uses the Wasserstein distance, is introduced to construct a spatial-temporal association matrix, allowing for the dynamic capturing of long-range correlations between nodes. The model's efficacy is validated through experiments on four publicly available traffic datasets. Results show that the proposed model consistently outperforms the best baselines in predictive accuracy. Furthermore, this study examines factors such as training data size, dimensionality, the number of attention heads, and the threshold of the spatial-temporal association matrix, and includes an ablation study to evaluate the model's overall performance.