Causal-Enhanced Spatio-Temporal Graph Neural Point Processes for Traffic Congestion Event Prediction

Abstract

Urban congestion remains a persistent challenge for cities, where disruptions often propagate across road networks in complex and poorly understood ways. While recent models such as the Spatio-Temporal Graph Neural Point Process (STGNPP) provide strong performance in congestion event prediction, they offer limited interpretability for understanding why and how congestion spreads. Conversely, Granger causality serves as a prominent approach for uncovering directional dependencies in traffic flow but struggles to scale to large, high-dimensional transportation networks. This gap highlights the need for a unified framework that combines both predictive accuracy and causal interpretability. This research proposes a Causal-Enhanced STGNPP framework that integrates congestion event prediction with scalable causal analysis. Building on the strengths of STGNPP and inspired by recent work such as the Spatio-Temporal Granger Causality Graph Neural Network (STGC-GNN) for traffic speed prediction, the proposed model introduces a Granger-based causal discovery module to identify statistically significant causal pathways of congestion propagation. Using a comprehensive traffic dataset from Xuancheng, this study investigates whether incorporating causal structure can improve event prediction performance and provide interpretable explanations for network-wide congestion dynamics. Given this motivation, the project aims to address the following question: Can causal analysis be effectively integrated into spatio-temporal neural models to enhance congestion prediction and uncover interpretable propagation mechanisms? To answer this, the study sets two quantitative objectives: (1) improving the congestion-event Neural Point Process (NPP) learning objective (evaluated as MAE𝑡 for the inter-event time between consecutive congestion events), and improving the duration modeling accuracy of congestion events (evaluated as MAE𝑑 for event duration); and (2) identifying statistically significant causal relationships among road segments, whose measurability and interpretation are evaluated via spatio-temporal Granger causality analysis. By bridging deep learning with causal inference, this work seeks to advance both methodological development and practical decision-making for real-world transportation systems.