Causal Impact Inference for Traffic Networks with Graph-Integrated Transfer Entropy

Abstract

<p>Causal graph is an indispensable part of causal analysis, which plays a considerable role in congestion and accident investigation, as it provides a more intuitive form of causal correlations in traffic networks and the causal impacts' propagation process. Previous studies have generated the causal graph based on information theory and model-based approaches. However, the lack of attention to region-wise mutual information and the ignorance of network topology hinder the comprehensive understanding of the causal relations. To address such limitations, this work provides a new perspective to achieve the region-wise causal inference of complex traffic network with enhanced verifiability and interpretability. Specifically, the trainable causal entropy algorithm is first applied to obtain the initial causal matrix, considering the original topology structure of the network. A causal-integrated Graph Convolutional Network (GCN) with attention mechanism is then designed to capture region-wise causal effect by information aggregation through spatial convolution. The initial causal graph is iteratively refined through the end-to-end framework and the improvement of traffic flow forecasting accuracy evaluated on real-world dataset using causal graph indicates the effectiveness of the proposed method in capturing causal relations. We performed comparisons between topology and causal graph, and the visualization of the asymmetric causal graph can provide the understanding of traffic dynamics and underlying causal impacts.<br></p>