The Research For Point-of-Interest Recommendation Based On User Check-in Behavior

With explosively development of mobile internet technology and emergence of Beidou Navigation Satellite System,users can set foot in almost all regions in the world,as a result of which a lot of multimedia information with the tag of geographic location was generated and location-based social networks were formed.In order to recommend intriguing locations to users,build user-friendly travel pattern,and rational distribution of social resources,location recommendation system has evolved into a hot research spot.The traditional location recommendation systems usually make use of methods of collaborative filtering or matrix factorization to calculate the predictive score of each location,according to which the user can be recommended.However,the methods above generally use one-hot check-in vectors to calculate user similarities,which do not handle the data sparsity well,and the recommendation performance tends to be poor as a result of untrusted similarity measures when the users have only a few check-ins.Even worse results on the cold start problem.By mining the context information such as geographical,temporal and social information over the user’s check-in data,the recommendation not only shows good personalization performance on different dimensions,but also problems of the sparsity and cold-start can be alleviated so that a better user experience can be improved.Firstly,this paper analyzes the spatial attribute and temporal attribute of users’ check-in data in location-based social networks,and summarizes spatial clustering phenomenon and the sequential transition feature of user’s check-in from the data analysis.Then a hybrid recommendation method is proposed,which mainly includes the following contributions:(1)In terms of spatial features,the hierarchical clustering and two-dimension kernel density estimation are combined together in order to prevent the problem that personalized points of interest are submerged popular ones.According to the Haversine distance,check-ins of users are hierarchically and spatially clustered,and multiple cluster centers that users visited frequently are automatically obtained.Then the two-dimensional kernel density estimation by latitude and longitude coordinates in each user cluster are averaged.(2)In aspects of time features,random walk on full network is achieved on basis of first-order location transfer matrix to solve the sparseness problem.Using the check-in sequence and time interval of each user,a transition probability matrix of location-to-location is constructed.Meanwhile,heterogeneous random walk on graph with restart is used to obtain the probabilities of steady state that user visits point-of-interest by passing the user’s preferences among whole graphs iteratively.(3)In terms of social factors,in order to ease the influence of dense check-ins of friends on user preference scores,the paper improved collaborative filtering and proposed a hybrid multi-context recommendation method.Using an improved social-based collaborative filtering approach,users can share interest with their friends.Finally,based on the geographical preference,temporal preference and social preference,a hybrid recommendation method is proposed to calculate the user’s probability of visiting point-of-interest,which is useful to solve the problem of cold start.The analysis of the final experimental results shows that the hybrid recommendation method proposed in this paper is superior to the commonly used benchmark methods in the precision and recall metric and to a certain extent,the problems of cold start and data sparsity problem are alleviated.