CL-DGCN: contrastive learning based deeper graph convolutional network for traffic flow data prediction

Accurate and efficient traffic prediction helps to establish multimodal transportation systems and improve the travelling experience in daily life. Currently the mainstream traffic prediction methods are implemented based on Graph Convolutional Network (GCN), superimposing GCN layers can obtain better prediction results, but excessive superimposition will lead to the over-smooth problem, this paper proposes CL-DGCN to overcome this problem, which obtains the representations of the features through contrastive learning, and uses the improved message aggregation function to overcome the over-smooth problem. In this study, the CL-DGCN model is experimented on four domestic and international open-source, real datasets (PEMSBAY, METR-LA, BEIJING and SZ-TAXI), and CL-DGCN achieves optimal or sub-optimal results in most time-step predictions, and reduces the composite error by more than 10 % compared to the baseline model, which well illustrates that the CL-DGCN model possesses more accurate prediction results.