Enhancing Spatial-Temporal Prediction Models with Dynamic Causal Graphs

Spatial-temporal prediction has become an important task in many applications, such as traffic forecasting. Due to the spatial-temporal nature of data, most state-of-the-art methods heavily depend on graph neural networks to model the inherent spatial relationships. However, most of them process the spatial data by applying the prior adjacency knowledge or learning a static adaptive adjacency matrix. Thus, their prediction performance is limited on dynamic situations where the spatial dependencies change w.r.t time. Furthermore, considering the stochastic training process, learning an adaptive adjacency matrix from scratch also makes it difficult for the neural network to achieve stable parameters and performance. To address the above challenges, this article proposes three practical extensions that incorporate dynamic causal knowledge into the training of graph convolution networks. We first analyze the dynamic causal graphs between traffic nodes with one dynamic causal discovery algorithm in each extended model. Subsequently, the spatial module employs dynamic causal graphs to reveal the evolving connections among nodes. Extensive experiments demonstrate that our method has successfully enhanced state-of-the-art traffic forecasting models on two benchmarks.