Research On User Check-in Characteristic Analysis And Recommendation Methods In Location-based Social Networks

In recent years,with the rapid development of location-based social networks(LBSNs),mobile terminals and cities,user’s check-in data has grown explosively,and it is affecting or changing people’s life at an unprecedented scale and speed.These check-ind data hide users’ personal preferences,activity tracks and patterns of life,etc.Therefore,the collection,mining and analysis of user check-in data have attracted extensive attention from academia and industry,aiming to better understand users’ mobile behavior,improve user experience and meet user needs.In addition to the trend of the growth of data scale,the users have increasingly strong demands for intelligent and personalized services.The improvement of users’ demands also reversedly promote the diversification of service modes and applications of various LBSNs software,and promotes the transformation of the research objectives and contents of the location recommendation systems.More and more research on location recommendation algorithms and applications arise,aiming to solve the problem of information overload,improve the recommendation accuracy,and meet the user’s personalized and intelligent service needs.Although machine learning and deep learning technologies make location recommendation develop both in theory and practice,due to the lack of deep mining of user behavior in complex context and data sparsity,there is still space for improvement in performance and quality of the recommendation system.Therefore,in the rapid development of the LBSNs and the demand for increasingly precise of personalized location recommendation,how to make full use of context information in modeling the behavior of users’ preferences,recommend locations or activities of interest to the users,is one of the key issues to improve recommendation performance and user experienceis.This thesis summarizs the existing researches and deficiencies in LBSNs from three aspects,i.e.point-of-interest recommendation,region-of-interest recommendation and activity recommendation.Focusing on the geographic features of users,we analyze and make empirical researches from three aspects using three LBSNs public datasets: POI recommendation model based on geographical and social information,ROI recommendation model based on geographical,categorical information and city blocks,activity recommendation based on spatiotemporal context.The main research works are as follows.(1)This research analyzes user’s check-in features from geographic,temporal,categorical and social aspects on three publicly LBSNs check-in datasets(i.e.Foursquare,Gowalla and Yelp).Statistical methods and spatial analysis methods were used to draws user’s activity preferences.Results show that user’s check-in behaviors are significantly affected by the geographic information,temporal information,categorical information and social information of the points of interest,and its closely related in temporal and spatial context.The results provide support for the design of the personalized location recommendation algorithm.(2)A point of interest recommendation method based on geographical and synthetic social influences(Ge SSo)is proposed.In view of the existing two-dimensional kernel density estimation model cannot accurately find out user’s preferences because of excessive bandwidth or outliers,a personalized two-dimensional kernel density estimation method is proposed.The model optimized the bandwidth calculation,and modeling two-dimensional check-in probability distribution for each user on the latitude and longitude.A comprehensive social impact model was proposed to fully assess the social information which considers the social tightness and social relations between friends.Finally,through the linear integration of user preferences,geographic influence and social influence,personalized POI recommendation is realized.Based on the idea of collaborative filtering,the Ge SSo model utilizes geographic and social information to improve the sparsity of user check-in data and the accuracy of the POI recommendation.(3)A region of interest recommendation method based on geographical,categorical context,and city block(GCCB)is proposed.The boundaries of user’s ROIs recommended by the existing method are arbitrary,which have no actual geographic sense and is not highly interpretable.In this paper,urban block data is taken as the preset partition result,and the user’s region check-in can be obtained based on the user’s POI check-in data.Then personalized modeling of spatial activity preference and category preference is carried out to study the importance of geographical influence and category influence in improving recommendation performance.Finally,a context framework was used to merge the spatial and categorical influence.This model takes into account the sparsity and bulkiness of the check-in data,and experiments on real datasets show that the method has high accuracy,which is of certain significance for the recommendation and prediction of the urban blocks that users are interested in.(4)User activity recommendation based on spatial and temporal context(STUAR)is proposed.To solve the problem of data sparsity in context-aware user activity modeling,this paper proposes to integrate user spatial activity preferences and temporal activity preferences by constructing user interest regions.In the spatial preference model,the ROIs of users are mined by constructing regular grid,and the spatial preferences of users are inferred based on the ROIs.In the temporal preference model,tensor decomposition method is used to infer users’ temporal preferences.Finally,based on the current context of the user,the spatial and temporal activity preference are fused,the potential relationship between the user,location,time and activity is mined,and the activity category that the user may be interested in is recommended.STUAR takes into account the sparsity of the check-in data,and experiments on real datasets show that the method has high accuracy,which is of significance for the recommendation and prediction of activities that user is interested in.