Research On Tensor Completion Algorithm Based On Common Structure Of Multi-source Data Sets

Traffic congestion and safety problems caused by the increase in cars test the level of urban traffic management,increase the carrying capacity of the transportation system,improve travel conditions,and provide smart decisions for future urban hotspot area mining and urban planning.This has become hotter research issues recently.Driving data is the prerequisite and basis for intelligent transportation system planning.A complete transportation network and vehicle operation data can provide high-quality transportation service information,and improve the comprehensive operation efficiency and intelligent management level of the transportation system.At present,many cities’ road networks have deployed traffic monitoring equipment to obtain real-time traffic data.However,due to problems such as collection equipment failures,data transmission and storage failures,the quality of real-time acquired data is usually not high and will exist.Problems such as missing data,errors,and redundancy make it difficult for smart traffic control and guidance to be carried out.In this research,with aiming at the problem of traffic data lackness,based on the common structural characteristics of multi-source data sets and the spatio-temporal multi-modal characteristics of the data,the tensor decomposition algorithm is used as the core of the complement model to recover the traffic data while supplementing the traffic data.Research the whole issue.Traffic data completion problem,it is of great significance for the preprocessing of the data to discover and obtain the data features with a high degree of correlation,which is significant to improving the accuracy of data recovery.Carrying out tensor modeling on the research problem,using the advantages of the multi-modal data table structure of the tensor model and the parallel processing of multi-dimensional data can not only improve the accuracy of data recovery,but also greatly improve the operating efficiency and performance of the model.Combining real GPS data and hotspot data in urban areas,we discuss the time and space dimensions granular modeling technology based on tensor decomposition to recover traffic data.In the traffic road network information recovery algorithm,when some data is lost,some models perform poorly.This paper proposes a regular tensor decomposition model with residual tensor under the common structure of multi-source data.The best model feature rank provides high-precision completion of road network data in the transportation field.The model algorithm was verified experimentally through the GPS data of the Kunming Traffic Operation Department and the urban hotspot POI data.It shows that the completion model of regular tensor decomposition with residual tensor can achieve a good data recovery accuracy for missing data in a certain interval or time segment when the optimal rank is selected appropriately.Compared with traditional machine learning algorithms and statistical learning models,it is not only faster and better,but also has higher data recovery accuracy and good robustness.This article focuses on the following aspects of the completion of traffic data based on tensor decomposition and tensor correlation analysis,:（1）The classical tensor decomposition model cp and Tucker are subjected to variant decomposition and optimized by alternating least squares.After optimization,it can obtain better recovery accuracy and more complete data extraction than the least squares method,especially when dealing with sparse tensors.The amount of information.（2）Improve the alternative optimization method of variant decomposition and introduce tensor regularization gradient alternation and decomposition correction.On this basis,combined with real traffic road network data proposed a tensor completion algorithm based on the common structure of multi-source data sets.This algorithm not only The regular term is introduced into the optimization of variant alternate decomposition for over-fitting correction,and the residual tensor is introduced to further improve the completion accuracy.（3）With multi-source traffic data sets through experimental comparative analysis of models such as KNN,MICE,Miss Forest and TDIM,it is shown that the algorithm in this paper selects different tensor local optimal ranks to optimize the interval time slice of the experimental data.Recovery accuracy.The model is not only faster than the traditional model,and avoids the explosion of dimensionality.Through the optimal rank selection,the recovery accuracy is higher than the above model,and it has good robustness.