Abstract
Traffic prediction is a critical component of intelligent transportationsystems, enabling applications such as congestion mitigation and accident riskprediction. While recent research has explored both graph-based and grid-basedapproaches, key limitations remain. Graph-based methods effectively capturenon-Euclidean spatial structures but often incur high computational overhead,limiting their practicality in large-scale systems. In contrast, grid-basedmethods, which primarily leverage Convolutional Neural Networks (CNNs), offergreater computational efficiency but struggle to model irregular spatialpatterns due to the fixed shape of their filters. Moreover, both approachesoften fail to account for inherent spatio-temporal heterogeneity, as theytypically apply a shared set of parameters across diverse regions and timeperiods. To address these challenges, we propose the Deformable DynamicConvolutional Network (DDCN), a novel CNN-based architecture that integratesboth deformable and dynamic convolution operations. The deformable layerintroduces learnable offsets to create flexible receptive fields that betteralign with spatial irregularities, while the dynamic layer generatesregion-specific filters, allowing the model to adapt to varying spatio-temporaltraffic patterns. By combining these two components, DDCN effectively capturesboth non-Euclidean spatial structures and spatio-temporal heterogeneity.Extensive experiments on four real-world traffic datasets demonstrate that DDCNachieves competitive predictive performance while significantly reducingcomputational costs, underscoring its potential for large-scale and real-timedeployment.