Tensor-evolving graph with temporal separation network for traffic flow forecasting

Abstract

Traffic flow forecasting requires joint modeling of temporal nonstationarity and diverse evolution patterns of traffic states under non-Euclidean road network constraints. Although recent spatiotemporal graph forecasting methods incorporate dynamic graph learning to alleviate the limitations of static topology, their spatial structure evolution is still largely driven by feature similarity or latent variables. An explicit temporal-prior-modulated mechanism for structure generation remains absent. Temporal modeling and spatial structure learning are often conducted independently at different levels, limiting the ability to capture periodically driven evolution of spatiotemporal dependencies. To address these limitations, a Tensor-Evolving Graph with Temporal Separation Network (TEG-TSNet) for traffic flow forecasting is proposed, which constructs a unified spatial prior via graph Laplacian spectral embedding. A node-level gating mechanism is then applied to enable differentiable temporal decoupling between trend and seasonal components. Next, a tensor-evolving graph encoder embeds time, source nodes, and target nodes into a unified multilinear tensor representation. This design dynamically generates structure-aware adjacency relations for different time slices and models time-varying spatial dependencies via diffusion graph convolution. During decoding, spatiotemporal attention fusion and a sparse expert routing mechanism are employed to strengthen representations of multimodal traffic states. Unlike prior studies that loosely combine temporal modeling and dynamic graph learning, the proposed framework introduces a time-conditioned structure generation paradigm, where temporal priors explicitly govern the evolution of spatial dependencies through a unified tensor formulation. This establishes a tightly coupled spatiotemporal modeling mechanism rather than a simple component-level integration. Extensive experiments on multiple real-world traffic datasets demonstrate stable and competitive forecasting performance and validate effectiveness and generalization. Related codes are available at https://github.com/Haku-zx/TEG-TSNet.