Research On Successive Point-of-Interest Recommendation Model Integrating Social Connection And Geographic Location

With the spread of mobile location technology,many online social networking applications have introduced location-based services.As a result,location-based social networks(LBSN)are booming,making them an integral part of people's lives.In location-based social services,users usually share their experiences or comments with friends in the form of check-ins,resulting in massive check-in data,which provides a basis for mining user's preferences.Successive point-of-interest recommendation is a very important application in location social network.It learns the user's preferences by analyzing the user's check-in data,thereby recommending the user's next POI.However,due to the limitation of check-in record and the massive number of points of interest,the user's check-in data is far sparser than that of the traditional recommendation system.In addition,user check-in data contains various contextual information,including social information,geographic information,comment information,etc.However,these data are composed of different structures and are heterogeneous.Effective fusion of heterogeneous information is also essential for successive point-of-interest recommendation.This paper designs a model of successive POI recommendation named Social-Geographical Long Short-Term Memory that integrates social connection and geographic information in order to address above challenges.The main research in this paper is as follows:Firstly,based on two public LBSN check-in datasets,this paper analyzes the user's behavior.Specifically,this paper analyzes the correlation between social factors,geographic factors,time factors in LBSN and user check-in behavior.The results show that the user's performance is significantly affected by the contextual information,which provides ideas for the design of successive point-of-interest recommendation model.Secondly,this paper designs a new feature representation method based on non-negative matrix factorization to model social connection and geographic information in order to alleviate the sparsity of data.Specifically,for social information,this paper constructs a feature matrix containing both explicit and implicit social relationships.For geographic information,this paper constructs an association matrix and a geographic neighbor matrix.Then it calculates social feature vectors and geographic feature vectors.through non-negative matrix decomposition.Thirdly,this paper designs a successive point-of-interest recommendation model that integrates social connection and geographic information.The model integrates the user's social feature vectors,the geographic feature vectors and the time feature into the LSTM as contextual information,and the neural network can learn the interaction between multiple features to integrate the heterogeneous information.This model sees the recommendation as a multi-classification problem.The output is the probability of user's check-in.Finally,the model sorts the check-in probabilities,and gives several points-of-interest with the highest probability as the recommendation result.This paper experiments with the proposed model based on two publicly available LBSN check-in datasets,Gowalla and Yelp.Firstly,the SGLSTM model has a notable progress in Recall@k compared to the mainstream successive point-of-interest recommendation model.Secondly,it analyzes in detail the impact of the three main factors of social,geographic and time on the recommendation results.Thirdly,the sensitivity analysis of the parameters is performed to demonstrate the effectiveness of the model.