基于因果干预与动态图学习的交通事故风险预测

PDF下载HTML

Title:

Traffic accident risk prediction based on causal intervention and dynamic graph learning

作者:

孙翊文¹, 张艺驰², 陈嘉骐³

1. 北京大学 人工智能研究院, 北京 100871;
2. 百度在线网络技术有限公司, 北京 100193;
3. 中央民族大学 信息工程学院, 北京 100081

Author(s):

SUN Yiwen¹, ZHANG Yichi², CHEN Jiaqi³

1. Institute for Artificial Intelligence, Peking University, Beijing 100871, China;
2. Baidu Online Network Technology Co., Ltd., Beijing 100193;
3. School of Information Engineering, Minzu University of China, Beijing 100081, China

关键词:

事故预防; 因果干预; 深度学习; 预测; 风险预测; 图神经网络; 时间序列; 长短期记忆网络; 交通控制; 卷积神经网络

Keywords:

accident prevention; causal intervention; deep learning; forecasting; risk prediction; graph neural networks; time series; long short-term memory; traffic control; convolutional neural networks

分类号:

TP391.4; U491.31

DOI:

10.11992/tis.202511031

摘要:

针对城市交通事故风险预测中存在的泛化能力差、时空相关性建模不足的问题，提出一种基于因果干预与动态图学习的交通事故风险预测网络。该网络在编码器-解码器框架中，设计了输入门模块以融合多周期事故信号与外部协变量；提出动态图学习模块，通过通道注意力机制自适应刻画节点间时变的空间依赖；并结合图卷积与门控循环单元联合建模时空演化规律。为进一步提升预测稳定性，引入转移桥模块，将历史时空状态映射至未来状态，以缓解多步预测中的误差累积问题；同时，采用基于前门准则的因果干预机制，从特征层面切断混杂因子路径，增强模型的鲁棒性与可迁移性。在NYC和Chicago两个公开数据集上的实验结果表明，CIDGNet在均方根误差、召回率和平均精度均值3项指标上均优于4个基线模型。消融实验进一步验证了动态图、输入门、转移桥等关键模块对性能提升的贡献，证明了因果干预与动态图学习在复杂时空预测任务中的有效性。

Abstract:

To address the limited generalization ability and inadequate spatiotemporal dependency modeling in urban traffic accident risk prediction, this paper proposes a traffic accident risk forecasting network based on causal intervention and dynamic graph learning. Within an encoder-decoder framework, an input-gating module is designed to fuse multi-period accident signals with external covariates. A dynamic graph learning module is further introduced to adaptively characterize time-varying spatial dependencies among nodes via a channel-attention mechanism, while graph convolution and gated recurrent units are jointly employed to model spatiotemporal evolution patterns. To further improve forecasting stability, a transition-bridge module is incorporated to map historical spatiotemporal states to future states, thereby mitigating error accumulation in multi-step prediction. Meanwhile, a front-door-criterion-based causal intervention mechanism is adopted to sever confounding paths at the feature level, enhancing model robustness and transferability. Experiments on two public datasets, NYC and Chicago, demonstrate that CIDGNet outperforms four baseline models in terms of root mean square error, recall, and mean average precision. Ablation studies further verify the contributions of key components-including the dynamic graph, input gate, and transition bridge-to performance gains, confirming the effectiveness of causal intervention and dynamic graph learning for complex spatiotemporal forecasting tasks.