Point-of-interest Recommendation In Location-based Social Network Considering Offline Interaction

The popularization of smart mobile devices and 4G technologies has brought tremendous development of large-scale location based social networks(LBSN)applications such as Dianping and Discover.At the same time,with the rapid development of the city,the number of Point-of-Interests has also increased.People are becoming accustomed to discovering,accessing,and evaluating Point-of-Interest through applications that are location-based social networks.Based on the massive accumulation of location-based social network data and the desire of application platforms to improve the user experience,the Point-of-Interests recommendation system has gradually attracted the attention of industry and academia.However,the Point-of-Interests recommendation face some new problems such as how to use user preferences,geographic location,social networks,and other contextual information in a comprehensive manner,How to solve the sparseness problem of user check-in data and How to handle the implicit user feedback.In response to these challenges,this thesis uses social relationships which consider the offline interaction behaviors of users to implement check-in matrix filling to mitigate data sparsity;integrates multiple data and improves geographic location impact modeling,and proposes a Point-of-Interests recommendation model considering offline interactions.For the user preferences in the model,a weighted matrix factorization algorithm with inline classification and a ranking-based algorithm with inline classification are proposed.The details are as follows:(1)Through the analysis of users’ social networks,it is proposed to use a social relationship that takes into account offline interactions to implement the filling of the check-in matrix,and to alleviate the problem of data sparseness through the similar preferences of friends.Experiments in two real datasets have shown that matrix filling is meaningful.(2)Based on the importance of geographic location impact in location-based social networks,the geographic impacts are modeled as non-parametric distributions on the individual user’s,and the use of adaptive kernel density estimation improves the geographical distribution ability of POI.Based on the heterogeneity of location-based social network data,and the use of social relations to achieve matrix filling,with integrating the influence of user preferences,classification and geographical location,a POI recommendation model considering offline interaction is proposed.(3)Based on the implicit of user feedback in the location-based social network,the user’s check-in number of POI is regarded as the user’s confidence.A weighted matrix factorization of POI recommendation algorithm that takes into account offline interactions is proposed,and an alternative least squares method is adopted to optimize learning;In order to improve the use of negative samples of the check-in data during the learning process,and avoid the prediction bias due to the square error during the learning of the scoring matrix,an considering offline interactions POI recommendation algorithm is proposed,and a random gradient descent algorithm is adopted to optimize learning.The experimental results in two real data sets show that the proposed methods improve the effectiveness of the POI recommendation and have practical significance on the application of location-based social networks.