A Multi-Graph Attentive Network for Traffic Speed Prediction and Diagnosis

Urban traffic speed prediction with high precision is the unremitting pursuit of intelligent transportation systems. The fundamental challenges of traffic speed prediction lie in the accurate modelling of the complex temporal and spatial correlations of transportation systems. Among all the methods, the hybrid “GNN + RNN” models have achieved state-of-the-art results. However, these methods still cannot address the following two challenges. First, in addition to the topology of road networks, the traffic speed could be affected by a variety of other factors, such as road functionality and weather. Second, in addition to predicting traffic speed, it is necessary to diagnose the causes of the prediction results. In this paper, we propose a multi-graph attentive network (MGAN), to predict and diagnose urban traffic speed. We create GNN model by using multiple graphs to encode the factors affecting them from various aspects. And we design a hierarchical attention mechanism to organize and pinpoint the fine-grained effects of different affecting factors for diagnosing the prediction results. The experimental results demonstrate that MGAN achieves state-of-the-art prediction performance on two real-world datasets, outperforming the strongest baseline by at least $5.94\%$ across three prediction horizons, and is able to intuitively diagnose the prediction results.