大数据处理中基于多任务学习的交通预测框架  

Traffic prediction framework based on multi task learning in big data processing

作　　者：殷正坤[1] 李鹏 YIN Zheng-kun;LI Peng(College of Economics,Trade and Information Technology,Changsha Vocational and Technical College,Changsha 410217,China;School of Information Science and Engineering,Hunan University of Chinese Medicine,Changsha 410208,China)

机构地区：[1]长沙职业技术学院经贸与信息技术学院,湖南长沙410217 [2]湖南中医药大学信息科学与工程学院,湖南长沙410208

出　　处：《计算机工程与设计》2022年第5期1443-1450,共8页Computer Engineering and Design

基　　金：国家社会科学基金重点基金项目(17AZD037);国家重点研发计划基金项目(2017YFC1703306);湖南省自然科学基金青年基金项目(2019JJ50453);湖南省教育厅科学应用研究基金项目(18C1857)。

摘　　要：现有的交通预测方法很少利用交通传感器之间的共性来做预测,很少有研究会为不同的交通情况建立不同的模型,影响了预测的精度。为此,提出一种基于多任务学习的交通预测框架(MTL-SA)。对于所有的交通传感器数据,采用基于非负矩阵分解的聚类来区分不同的交通情况;对于每种特定的交通情况,基于组lasso和l_(2)范数惩罚的多任务学习来建模交通预测问题;采用能保证收敛率的FISTA算法进行问题求解。基于大量实际交通传感器数据的实验结果表明,所提MTL-SA框架始终比每个传感器单独应用MTL的预测结果要好。在不同交通情况下,所提方法在短期预测和长期预测中的性能比其它典型的交通预测方法分别高出18%和30%。The existing traffic prediction methods seldom use the common features of traffic sensors to make predictions,and few studies build different models for different traffic situations,which affects the accuracy of prediction.To this end,a multi task learning based traffic prediction framework(MTL-SA)was proposed.For all traffic sensor data,clustering based on non-negative matrix decomposition was used to distinguish different traffic situations.For each specific traffic situation,the traffic prediction problem was modeled based on group lasso and norm penalty.The problem was solved using the FISTA algorithm that guaranteed the convergence rate.Experimental results based on a large number of real traffic sensor data show that the prediction results of the proposed MTL-SA framework is always better than that of each sensor using MTL alone.Under different traffic conditions,the performance of proposed method in short-term prediction and long-term prediction is 18%and 30%respectively higher than that of other typical traffic prediction methods.

关 键 词：交通预测 多任务学习 非负矩阵分解 聚类 FISTA算法 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]