• Propose an Adaptive Graph-Aware Clustering (AGAC) traffic flow forecasting model to address spatio-temporal heterogeneity. • Introduce an adaptive periodic embedding fusion mechanism to dynamically model multi-scale temporal variations. • Construct a similarity-driven multilayer graph to organize nodes into hierarchical structures balancing local and global dependencies. • Develop a cross-layer contrastive learning strategy with adaptive pair selection to selectively align heterogeneous spatio-temporal representations. Traffic flow forecasting plays a crucial role in intelligent transportation systems. Although existing research has made significant progress in modeling spatio-temporal correlations, they often lack capturing the dynamic characteristics induced by changes in traffic patterns. Moreover, it is still difficult to model the inherent heterogeneity across different temporal scales and spatial regions, as they typically treat the road network as a homogeneous entity with similar traffic fluctuations. To address these challenges, we propose an Adaptive Graph-Aware Clustering (AGAC) traffic flow forecasting model, which captures and adapts to evolving traffic patterns by constructing a multilayer graph and employing cross-layer contrastive learning. Specifically, our approach captures spatio-temporal dependencies via convolutional operations and utilizes an adaptive gating mechanism to dynamically adjust periodic embeddings, thereby capturing traffic flow trends at different temporal scales. Node similarities are calculated to construct a multilayer graph and iteratively refine the graph structure through cross-layer contrastive learning, thus enhancing the model’s capacity to adapt to continuously evolving heterogeneous traffic patterns. Experiments on five benchmark datasets demonstrate that the performance of AGAC consistently outperforms that of various state-of-the-art baselines. The source code is available at https://github.com/kaikai-kks/AGAC .
Zhang et al. (Mon,) studied this question.