Research On Trajectory Data Recovery Based On Deep Learning

With the wide application of positioning devices and video capture devices,the acquisition of spatiotemporal trajectory data becomes more and more convenient.For example,people share their location with friends on social networking sites,which generates a large amount of check-in data.Check-in data and vehicle traffic data are typical space-time trajectory data.They all contain object,location and time attributes.These data contain tremendous value and there have been a lot of achievements in the areas of urban computing,path planning,and location prediction.The trajectory recommendation or travel planning is one of the most important tasks in data processing tasks faced by a location based social network(LBSN).Traditional solutions usually use historical traces of users with many manual screening features.These features include the popularity of points of interest,feature values extracted from reviews,pictures,and texts.These features are used to learn the travel preferences of these users.This builds a travel trajectory recommendation model.This thesis propose an improved Monte Carlo Expectation(MCE)method based on the Monte Carlo Expectation(hereinafter referred to as MCE).Using the Monte Carlo search tree algorithm,we use the popularity and time characteristics of interest points to customize reward functions.The trajectory recommendation was conducted.In the public data set,compared with the baseline method using popularity and time characteristics,the MCE queuing time(MQI)was optimal and the recommendation correctness(F1 score)was optimal.Because features are built based on prior knowledge or instance analysis,this does not apply to a variety of LBSN websites.For the trajectory data that only has the user's check-in to generate POI points,we cannot rely on the help of feature engineering.In this case,the trajectory recommendation model cannot be established through the feature engineering scheme.This thesis propose a Trip Recommendation methodology via trajectory Encoder and Decoder(called TRED)-a novel end-to-end tour planning approach that works in a semi-supervised learning manner without feature engineering.Specifically,TRED encodes historical trajectories into vectors in order to learn intrinsic characteristics of POIs and the transition patterns among POIs,which can be used to recommend a trip given the desired start and end point.Incorporating historical attention mechanism in our proposed sequence-to-sequence model yields better solution(s)to trip recommendation problem,as illustrated by our experiments,conducted on several public LBSN datasets demonstrating that TRED outperforms state-of-the-art methods in terms of both recommendation correctness(F1 scores)and visiting orderness(pairs-F1 score).