Research Of Point-of-intereset Recommendation In Location Based Social Network

With the development of mobile Internet,LBSN(Location-Based Social Network)becomes popular.Users' behavior data in the LBSN can reflect the users' preference to the related POI(Point-Of-Interest).Actually,the POI recommender system is devoted to mining the users' s behavior data,thus to learn the preference model and the decision process of users,so as to predict and recommend POIs to users.POI recommendation is not only beneficial for LBSN service providers who aims at advertising to the targeted users,but also helps users to discover POIs,thus POI recommendation is of great value to be academiclly researched and applied in industry.In this paper,we focus on the POI recommendation algorithm which mainly achieve improvement by modeling the preference model and the decision process of users with the check-in data including geographical,social and temporal information.The main contributions of this paper are as follows:(1)We give an analysis of users' behavior pattern and construct the corresponding feature modeling method.In this paper,we find that the user is other-similar and self-smilar from the social and temporal data analysis,and correspondingly construct the feature of our model.(2)We propose the preference model which fused of improved factorization machines and sparse auto encoder.Firstly,we learn users' preference model based on the improved factorization machines with social and temporal regularization terms.Factorization machines is a model which can learn well even with sparse training data,since it can learn the relevance between features by decomposing the feature into latent feature vector.Thus,the factorization machines can learn the original features and the extended fusion features at the same time.Further,we propose to use autoecoder to do feature compression based on the neural network.Autoencoder can be regarded as a method of high dimensional features fusion which can extract the main information of features out of noises.What's more,the compressed features are used to model the preference of users which can be learned well since less parameters are needed.(3)We propose the POI recommendation algorithm which is intergrated by the users' preference model and the geographical decision model.We finally propose the POI recommendation algorithm based on the users' preference model and the power-law distribution of users' checkin probability in distance.POI recommender system is different from classic recommendation system since the decision of visiting a POI is both related to users' preference and other decision factors.In this paper,we mainly consider the decision factor of geographical distance,and thus obtain the final POI recommdation algorithm.(4)We conduct the experiments based on the common evaluation dataset,and we also implement a POI recommender system.In this paper,we will firstly introduce the research background and reviewed literatures,and analyze the existing problems.Then,we will detailedly introduce our proposed POI recommendation algorithm.Further,we will introduce the expriments and give an analysis of the result.Moreover,we will introduce the demo system based on POI recommendation.