Research On Key Techniques Of User Visit Location Prediction Based On Geo-Social Networking Data

With the popularity of smart mobile terminals and advance in wireless communication and positioning technologies,Geo-Social Networks(GSNs),which combine location awareness and social service functions,are becoming increasingly prevalent and have been an important link between “Online” and “Offline” business.A large amount of user trajectory data has been accu-mulated during the use of GSNs.This kind of data is extremely rich in temporal-spatial semantic information,which brings new opportunities for exploring user behavior patterns in GSNs and predicting user future consumption trends.Research on user visit location prediction based on Geo-social networking data aims to predict users' future visit destinations through mining their historical trajectories,which is conducive to have insight into users' consumer intention to produce location-based service recommendation,so as to enhance user participation.It is con-ducive to businesses to formulate personalized promotion strategies to guide user consumption behavior,so as to promote regional economic development.It is also conducive to urban man-agers to improve the utilization rate of public transport,so as to improve urban infrastructure construction.It is an effective means to achieve win-win situation among users,businesses and cities,and therefore has gained extensive and in-depth attention from academia and industry.The essence of user visit location prediction based on Geo-social networking data is to comprehensively analyze the location data carried by user check-ins,comments or other behav-ioral records generated in GSNs,to mine various user mobility patterns and personal preferences hidden behind the data,and finally infer the location where a given user would visit in the future.Currently,considerable progress has been made in the research of user visit location prediction at home and abroad,which brings inspiration and reference for the development of this disserta-tion.However,there are still deficiencies in current studies.Firstly,most of the related studies aim to solve the problem of user next visit location prediction,little attention is put on given-time user visit location prediction.Secondly,many studies often incorporate different contextual fac-tors through a linear fusion strategy,which is difficult to depict the role of multiple factors in governing user visiting behaviors.Thirdly,most studies only involve single-modal user check-in data for user visit preference modeling,the exploration toward multi-modal user generated data is still extremely scarce.Moreover,due to the uniqueness of Geo-social networking data,current research generally faces following troubles: data sparsity,cold-start,context awareness and user interest drift.In order to overcome the above deficiencies and challenges,this dissertation firstly con-ducts a survey on related works,and successively carry out the following studies step by step to-ward the fine-grained user visit location prediction problem at the point-of-interest(POI)level:(1)Toward anytime user visit location prediction problem,the multi-feature fusion based loca-tion prediction approach is proposed.Through systematically analyzing the mobility pat-terns in user trajectories,we extract multiple features from temporal periodicity,global pop-ularity,personal preference,as well as social influence.Then we design a scoring model and a classification model to effectively fuse these features,after which we can produce the prediction result following the “classifying first,ranking next” strategy.Experimental results over multiple real-world GSN datasets show that the proposed approach performs desirably on various evaluation metrics.(2)To solve user next visit location prediction problem and anytime user visit location pre-diction problem at the same time,a POI embedding based location prediction approach is proposed.In the first step,we design a random walk method to sample POI sequences from the constructed weighted POI graph.Then,we apply the word embedding model to train POI embeddings in the low-dimensional vector space.Next,we aggregate user trajectory to calculate user visit preference based on the exponential time decay.In the end,we produce the prediction result by computing the visit probability of a given user toward a candidate location.Experimental results over real-world GSN datasets verify the effectiveness of the proposed approach on both location prediction tasks.(3)To solve anytime user visit location prediction problem,considering the information associ-ation and complementation in multi-modal GSN data,a two-stage user location prediction approach based on hierarchical temporal-spatial preference is proposed.In the first stage,we learn time-sensitive user hidden representation through hierarchical attention mecha-nism.In the second stage,we design a time-sensitive scoring model to derive the visit probability of a given user toward a candidate location,so as to produce the prediction re-sult.Experimental results over multiple real-world GSN datasets show that the proposed approach has desirable predictive performance on various evaluation metrics and outper-forms many comparison methods.