Adjacent-DBSCAN Enhanced Time-Varying Multi-graph Convolution Network for Traffic Flow Prediction

Abstract

Accurately predicting traffic flow is very challenging since the traffic flow is collected from hundreds of sensor nodes, which are affected by multiple factors. The existing methods based on graph convolution focus on unified spatial-temporal feature extraction for all nodes, while ignoring the differences between nodes and the correlation between internal and external features. To overcome these shortcomings, we propose an adjacent-DBSCAN enhanced time-varying multi-graph convolution network (ADETMCN) for traffic flow prediction. First, we use the adjacency matrix to optimize the DBSCAN clustering algorithm and divide the raw traffic nodes into multiple highly correlated node clusters. Then, the multi-graph spatial feature, composed of the adjacency graph feature, the correlation graph feature, and the travel intention graph feature, automatically explores the internal and external spatial features of nodes. Next, the time-variant feature, composed of the minute periodic feature, the hour periodic feature, and the daily periodic feature, automatically extracts short-term and long-term spatial-temporal correlation by fusing the multi-graph spatial features. Extensive experiments on five real-world datasets show that the performance of the proposed model is superior to the most advanced methods.