Missing Data Imputation For Travel Time Data Based On Joint Matrix Factorization

Traffic congestion is a common problem in urban development,and it is also a difficult problem for all countries in the world.It has an inseparable relationship with people's daily life and urban development,so the research in the field of transportation has aroused widespread attention from researchers.In order to obtain real-time traffic situation,many cities around the world have deployed traffic monitoring equipment and GPS-based services on the road network.A large amount of traffic data makes research in the field of transportation possible,which is of great significance to solve the problem of urban traffic congestion.For example,accurately predicting traffic conditions in the road network can effectively avoid traffic congestion.However,due to some uncontrollable factors,such as equipment or communication failure,traffic data missing is pervasive and inevitable,which becomes a notable obstacle to the research in the field of transportation.Therefore,our research theme is traffic data imputation.There are many indicators to describe traffic conditions,such as traffic flow,travel time,congestion degree,etc.In this paper,we take the relative travel time as an example to solve the problem of missing traffic data imputation.By data analysis,we find that the relative travel time has three characteristics,including road similarity,periodicity and temporal coherence.For the traffic data imputation task,the existing methods mainly include:(1)the average historical traffic data,for example,temporal-neighborhood method,ARIMA and ARIMA-based imputation models;(2)modeling traffic patterns of similar road segments,such as KNN and PPCA;(3)utilizing deep neural networks,for example,using MLP,CNN or other neural networks to predict missing values.However,these methods do not fully exploit and utilize the spatial and temporal characteristics of traffic data in data imputation.Data sparsity is a difficult problem in our research.Therefore,we design a joint matrix factorization module and propose a novel travel time completion model,TIM for short,based on the characteristics of relative travel time data.TIM models the three characteristics of data to estimate missing values.It consists of three steps:(1)designing the organization form of model input.We organize relative travel time data into the form of three-dimensional tensor,which is used as the input of TIM.This data organization form maintains the natural characteristics of traffic data;(2)using a joint matrix factorization module to model the periodicity and road similarity;(3)modeling the temporal coherence.We acquire local constraints and add it to the joint matrix factorization module,so that the three characteristics of the data can work cooperatively.Therefore,TIM not only ensures the original nature of the data,but also makes full use of the spatio-temporal information of the relative travel time.To evaluate our model,we conduct a series of experiments on a real traffic dataset.The experimental results indicate that considering three characteristics of data can improve the performance of the model.Moreover,TIM performs better than the baselines for the task of relative travel time imputation.