TEg-STGNN: Learning Time-Evolving Heterogeneity in Spatial-Temporal Traffic Networks

Spatial-Temporal Graph Neural Networks (STGNNs) have become a dominant paradigm for traffic modeling due to their ability to capture spatial and temporal dependencies in traffic networks. However, existing methods typically rely on a uniform representation for modeling temporal dependencies, which fails to capture the evolutionary heterogeneity of traffic dynamics under varying scenarios. Consequently, despite achieving competitive overall results, their performance is often compromised in critical scenarios. Specifically, traffic dynamics under different scenarios manifest as heterogeneous temporal patterns, which are generally categorized into short-term fluctuations with quick dissipation and long-term trends that evolve gradually. Effectively representing these temporal patterns while mitigating their interference is crucial for enhancing model generalization across diverse scenarios. To bridge this gap, we propose a Time-Evolving Spatial-Temporal Graph Neural Network (TEg-STGNN), a novel framework that incorporates the time-evolving heterogeneity of traffic dynamics. Our approach introduces a dynamic graph construction module that encodes scenario-specific patterns, along with a dual-pathway architecture designed to disentangle short-term disturbances from long-term evolutionary trends. A learnable gating network dynamically fuses these disentangled representations, ensuring robust prediction performance across diverse traffic scenarios. Comprehensive experiments on real-world datasets show that TEg-STGNN excels in predicting specific congestion scenarios and maintains strong generalization across a wide range of traffic conditions.