Interpreting Urban Regional Traffic Congestion Events via Textual Descriptions

Understanding urban regional traffic congestion contributes to the optimization of signal timing optimization, route planning, and emergency rescue. Current studies concentrate on interpretability not too much. To address this issue, we propose a post-hoc explainability framework that enhances the interpretability of Spatio-Temporal Graph Neural Networks (STGNNs) in traffic congestion forecasting. The framework leverages the Monte Carlo Tree Search (MCTS) algorithm to extract key subgraphs from the prediction graph and transforms the interpretation of regional traffic forecasts into human-understandable natural language descriptions. We validate the framework using 16 weeks of minute-level traffic flow data on a 6×7 grid road network in Shanghai simulated on SUMO. Results show that when predictions are made using explanatory subgraphs rather than their complements, the explainer model's MAE drops from 0.836 to 0.225 - a 73.1% reduction in relative error. Moreover, the average sparsity of the extracted subgraphs exceeds 96%.