DTG-LKNet: dual spatio-temporal graphs and large-kernel convolutions network for traffic prediction

Accurate traffic flow prediction is central to Intelligent Transportation Systems yet remains difficult due to non-Euclidean spatial structure, long-range propagation, and time-varying delays. However, existing deep learning methods have key limitations, including reliance on fixed temporal partitioning, small-kernel locality that restricts receptive fields, and graph constructions that lack long-range, dynamic correlations. This work presents dual spatio-temporal graph and large-kernel convolutionnetwork (DTG-LKNet), a Transformer-based architecture that packages a large-kernel convolution module with a dual spatio-temporal graph in a unified framework. On the temporal side, Deformable Patch Sampling learns sampling offsets around salient timestamps, and large kernels expand the effective receptive field without deep dilation stacks. On the spatial side, DTG-LKNet fuses a functional-similarity graph with the physical road network topology to represent long-range correlations. Comprehensive experiments on three large-scale benchmarks demonstrate consistent state-of-the-art performance against strong baselines. This article further visualizes the effective receptive field, validating the effectiveness of the large-kernel design.