User Preference And Embedding Enhanced POI Recommendation In LBSN

With continuous progress of mobile Internet and mobile device GPS,social networks and location-based services gradually merge to form a location-based social network,namely LBSN.The explosive growth of mobile social data and various spatio-temporal contexts on LBSN have caused users to face information overload and difficult choices.It is difficult to quickly locate their points of interest,so POI(Points-Of-Interest)Recommendation came into being.It aims to solve information overload,mine users' personalized preferences,and recommend POIs.However,POI recommendation still meets challenges such as various spatio-temporal contexts,the sparsity of the user-POI check-in matrix and nonlinear interactive modeling between users and POIs,and it is difficult to be effectively solved.With the rapid development of deep learning technology,more and more deep learning technology has been applied to POI recommendation and achieved obvious results.Therefore,based on deep learning technology,this paper has carried out POI recommendation model research on user preference and embedding enhancement in LBSN.The main work completed are as follows.(1)In the user's preference ehanced and embedding representation,most of the current research fusion LBSN context into models directly.If it cannot be effectively mined and extracted,it will cause the increase of the data dimension and cause the original sparse data to be sparser.Besides,mostly traditional collaborative filtering models based on matrix factorization are difficult to mine nonlinear interactions between users and POIs.In response to the above problems,a POI recommendation model based on users' preference of check-in and sociality,PSC-SMLP(Preference enhanced Spectral Clustering and Spectral enhanced Multi-Layer Perceptron),was proposed.It extracts user preferences by mining the spatial characteristics of the check-in information and social information in LBSN.Moreover,it combined spectral theory and neural network to build an embedding layer,mined nonlinear interactions between users and POIs,learned user's personalized preferences deeply and improve recommendation quality.(2)Furthermore,considering that POIs in users' check-in sequence in LBSN are often highly correlated,so mining the spatio-temporal context information of POIs through users' check-in sequence can model embedding representation of POIs better.Therefore,based on PSC-SMLP,this paper proposed a POI recommendation model based on the spatio-temporal characteristics of user check-in sequence,namely PSC-TTGR(Transformer enhanced Temporal POI Embedding and Geo Preference Ranking Factorization Machine).On the basis of PSC capturing users' preferences,PSC-TTGR learned embeding representation of POIs with spatiotemporal characteristics by fusing temporal and geospatial information.Then,it sorts preferences between POIs,so that a richer preference relationship can be constructed to train the ranking model.Finally,a candidate recommendation list is generated for the target user.(3)Gowalla,Yelp are selected to carry out experiment of the above model.Experimental results show that PSC-SMLP and PSC-TTGR both have better performance than baselines in terms of accuracy,recall and other evaluation indicators.PSC-TTGR also has better performance than PSC-SMLP.