基于超图的多任务注意力交通流预测模型  

A Multitask Attention Traffic Flow Prediction Model Based on Hypergraph

作　　者：韩浩东 陈盛双[1] 谢良 HAN Hao-dong;CHE Sheng-shuang;XIE Liang(School of Science,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学理学院,武汉430070

出　　处：《武汉理工大学学报》2023年第10期150-159,共10页Journal of Wuhan University of Technology

基　　金：广东省自然科学基金(2020A1515011208)。

摘　　要：针对现有的交通流预测模型中的不足,作者提出了基于超图的多任务注意力交通流预测模型HGAT-MTAN,改进了传统的图卷积神经网络。该模型首先利用循环神经网络和超图注意力网络提取多任务的共享特征;接着,为多个交通流预测任务分配私有特征;然后,通过引入多任务注意力机制,在不同任务之间实现正向知识转移;最后,为每个任务设计预测层,得到多个预测结果。将HGAT-MTAN模型与现有的多种交通流预测模型进行对照实验,结果表明,在不同城市的交通流数据集上HGAT-MTAN都取得了优于对照方法的预测性能。消融实验也进一步验证了模型组件的有效性。因此,HGAT-MTAN模型可以为交通流预测提供一种有效的解决方案,并提高交通流数据分析的准确性和可靠性。This paper proposes a hypergraph-based multi-task attention network model,HGAT-MTAN,for traffic flow prediction to address the shortcomings of existing traffic prediction models,which improves the traditional graph convolutional neural network.The model first uses a recurrent neural network and a hypergraph attention network to extract shared features for multiple tasks.Then,private features are assigned to multiple traffic flow prediction tasks.Next,the multi-task attention mechanism is introduced to achieve forward knowledge transfer between different tasks.Finally,a prediction layer is designed for each task to obtain multiple prediction results.HGAT-MTAN model is compared with existing traffic flow prediction models on different traffic flow datasets in various cities,and the results show that HGAT-MTAN performs better than the compared methods in terms of prediction performance.The ablation experiment further verifies the effectiveness of the model components.Therefore,HGAT-MTAN can provide an effective solution for traffic flow prediction,improving the accuracy and reliability of traffic flow data analysis.

关 键 词：交通流预测 超图神经网络 多任务学习 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]