Multi-Attention Based Spatial-Temporal Graph Convolution Networks for Traffic Flow Forecasting
Jun Hu, Liyin Chen
Abstract
Traffic forecasting is a great challenge to effectively extract complex spatio-temporal patterns due to the dynamic and nonlinear spatio-temporal relationships of traffic flow as well as many other constantly changing factors. A spatial-temporal graph convolution network (MASTGCN) based on multi-attention mechanism is proposed to predict long-term traffic conditions of different locations on the road network in this paper. MASTGCN consists of several independent spatial-temporal blocks and a fully-connected layer. More specifically, each block consists of two major parts: 1) Two gate-fused attention mechanisms to model spatio-temporal relationships in traffic data; 2) The spatial-temporal convolution that applies graph convolutions and customary commonplace convolutions to describe spatial and temporal features simultaneously. Our experiments on two real-world datasets demonstrate that our MASTGCN is superior to the existing state-of-the-art baselines by a significant margin.