Traffic prediction by graph transformer embedded with subgraphs

Abstract

Rapid urbanization and population growth raise significant challenges in modern traffic management, where traffic prediction is essential for intelligent transportation systems. Recent proliferation of graph neural networks also could not give us the satisfactory solution, because predicting traffic flow requires effective modeling of complex spatial correlations and temporal dependencies among sensors. In this paper, we propose a novel graph transformer that mitigates the spatial and temporal heterogeneity simultaneously. Graph partitioning to capture spatial heterogeneity induces the nodes grouped with similar contextual properties. The proposed transformer effectively handles long-term temporal dependencies, and combines subgraph embeddings to represent the correlation of global patterns. Experimental results on four traffic prediction benchmark datasets with high spatial dependencies show that the proposed method produces a 12.33%p performance improvement against the 14 state-of-the-art methods. Especially, it exhibits excellent performance in 60-minute predictions, and training times are comparable to the competitive methods.