Relation-diffusion augmented network-wide traffic state estimation under sparse sensor deployment

Abstract

Network-wide traffic state estimation under sparse sensor deployment aims to infer traffic conditions at unobserved road segments from limited sensor observations. Existing diffusion graph convolutional methods mainly propagate node features over physical topology or adaptive graphs, but the adaptive relationships involving unobserved nodes are often weakly supported because their traffic features are missing, zero-padded, or represented only by initialized embeddings. To address this limitation, this study proposes the Relation-Augmented Diffusion Graph Convolution Network (RADGCN), a relation-aware adaptive diffusion framework for traffic state estimation under sparse observations. RADGCN first learns functional relationships among observed nodes using reliable temporal traffic features and graph-aware attention, and then propagates these relationships to the full network through learnable bilateral diffusion kernels. The resulting relation matrix is incorporated into diffusion graph convolution as an additional transition operator, allowing traffic states to be estimated through both physical topology and learned functional dependencies. An auxiliary edge prediction task is further introduced to regularize node embeddings under inductive K-order neighborhood subgraph training. Experiments across three benchmark datasets show that RADGCN consistently outperforms state-of-the-art baselines across different sparse-sensor settings, achieving average improvements of 2.97%–11.82% in MAE/RMSE/MAPE. Additional ablation results verify the effectiveness of relation diffusion, relation-augmented feature propagation, and K-order subgraph sampling. Further experiments also validated RADGCN's generalizability across datasets and robustness to sensor sparsity.