GADEF-Net: A heterogeneity-aware dual-graph framework for robust multimodal traffic forecasting

Abstract

Effective traffic forecasting in Intelligent Transportation Systems (ITS) requires the coordinated use of heterogeneous sensing streams, yet many existing graph-based models still process all modalities with a uniform encoder. This design overlooks the fact that macroscopic traffic states and auxiliary contextual signals may exhibit different spatiotemporal propagation patterns. To address this issue, we propose the Adaptive Gated Dual-Graph Network (GADEF-Net), a heterogeneity-aware multimodal forecasting framework. GADEF-Net adopts a dual-branch architecture in which an attention-based branch captures the global temporal evolution of target states, while a diffusion-convolution branch models the localized propagation of auxiliary contexts. The two representations are integrated through an Adaptive Gated Fusion (AGF) module that dynamically adjusts fusion weights according to real-time context. Experiments on three real-world datasets show that GADEF-Net achieves strong overall forecasting performance, obtaining the best results on Daegu-Urban and PeMS08 and remaining competitive on PeMS-BAY. These results suggest that explicitly modeling heterogeneous propagation mechanisms can improve multimodal traffic forecasting, particularly in more stochastic settings and at longer prediction horizons, while also introducing additional computational cost for practical deployment.