LI, Tianhao (The University of Hong Kong), Zhao, Zhan (The University of Hong Kong), Liu, Xintian (The University of Hong Kong)

Adaptive Fusion of Decomposed Traffic Components: A Heterogenized Spatio-Temporal Attention for Traffic Forecasting

Scheduled for presentation during the Regular Session "S30a-Intelligent Modeling and Prediction of Traffic Dynamics" (TH-LM-T30), Thursday, November 20, 2025, 10:30−10:50, Gold Coast

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Accurate traffic flow forecasting remains challenging due to the heterogeneous spatiotemporal patterns and varying traffic components. To address this, we propose DCSTFnet (where DCSTF denotes decomposition, componential-spatial-temporal attention, and fusion), a novel traffic flow forecasting framework that decomposes traffic data to enhance predictive accuracy. The model first leverages a vehicle-attribute classifier and Discrete Wavelet Transform to split traffic time series into components representing distinct temporal dynamics and behavioral patterns. Each component is enriched with cross-component information and processed via a multi-channel spatiotemporal encoder. By integrating both dynamic local neighborhoods and globally representative nodes, DCSTFnet enables both local and global attention-based message passing among nodes. An adaptive fusion module with multi-level supervision integrates component-wise predictions into final forecasts. Experiments on real-world traffic datasets show that DCSTFnet consistently outperforms baselines in both accuracy and robustness, particularly under varying flow intensities and prediction horizons.