3D Graph Convolutional Networks with Temporal Graphs: A Spatial Information Free Framework For Traffic Forecasting

Spatio-temporal prediction plays an important role in many application areasespecially in traffic domain. However, due to complicated spatio-temporaldependency and high non-linear dynamics in road networks, traffic predictiontask is still challenging. Existing works either exhibit heavy training cost orfail to accurately capture the spatio-temporal patterns, also ignore thecorrelation between distant roads that share the similar patterns. In thispaper, we propose a novel deep learning framework to overcome these issues: 3DTemporal Graph Convolutional Networks (3D-TGCN). Two novel components of ourmodel are introduced. (1) Instead of constructing the road graph based onspatial information, we learn it by comparing the similarity between timeseries for each road, thus providing a spatial information free framework. (2)We propose an original 3D graph convolution model to model the spatio-temporaldata more accurately. Empirical results show that 3D-TGCN could outperformstate-of-the-art baselines.