Research On Multi-factor LBSN Recommendation Algorithm Based On Model And Semantic Backgroun

LBSN(location based social networks)is a complex heterogeneous information network that provides location services for users.When a user has visited a point of interest,he or she will check in at that point.The LBSN recommendation algorithm is dedicated to recommending the points of interest that users have not visited but may visit,so as to provide users with better services.Effective discovery of LBSN complex network structure is helpful to improve the recommendation accuracy,but the existing multiple community discovery methods cannot well integrate various semantic information in LBSN to construct the network structure of LBSN.This paper proposes a multi-factor joint clustering recommendation method based on non-negative matrix factorization,by establishing and solving the objective function based on non-negative matrix tri-factorization,the method can obtain the interrelated and internally close user clusters and interest clusters,and recommend points of interest and friends to users based on the idea of collaborative filtering.in order to integrate more semantic information,the secondorder meta-path method is used to calculate the similarity between users and points of interest as a constraint term.Sequential pattern appears when users check in points of interest,In order to effectively alleviate the data sparsity and consider the influence of the order on the recommendation effect during the recommendation process,we construct a location-location transfer matrix based on all the check-in records,the user’s preference value for a certain place is calculated by additive Markov chain,and compute the matrix.In the process of nonnegative matrix factorization,the initial value has strong influence on the operation result.Therefore,this paper proposes an initialization method based on spectral clustering to provide good initial values for prediction model construction.This paper conducted experiments on Gowalla and Yelp,real LBSN data sets.Compared with other algorithms,the proposed method has better recommendation effect.The effectiveness of the proposed initialization method is demonstrated by comparing the decreasing speed of the objective function value with different initial values.In addition,the paper also studied the influence of different factors and different parameter values in the experiment on the recommendation effect.