When Dynamic Causality Comes to Graph-Temporal Neural Network

Spatial-temporal data forecasting is a core task in many applications, and traffic forecasting is a typical example. Researchers have proposed various methods to explore spatial and temporal characteristics to improve forecasting accuracy, including the recently emerging graph convolution networks. However, most of them only consider the road network's prior knowledge or the graph's static characteristics and thus ignore the dynamic and deeper information hidden in the data. This paper presents a novel module based on dynamic causality analysis and graph convolution to integrate statistical theories and deep learning for better capturing spatial dependencies. Then we apply the module to two specific models. In each model, we introduce the causality adjacency matrix computed by the proposed algorithm into the conventional graph convolution network to reveal the dynamic correlations between nodes in the road network. The temporal neural network is then applied to extract temporal correlations. Extensive experiments demonstrate the superiority of our method, which achieves state-of-the-art prediction accuarcy on two public transportation data sets.