Research On Integrating Geographic Information Into A Matrix Factorization Model For Point-of-Interest Recommendation

With the widespread popularity of smartphones and the expansion of mobile communication services,Location-based social networks(LBSN)has ushered in rapid development.Point-of-interest(POI)recommendation is a personalized service popular in LBSN.POI recommendation determines the POIs that the user may be interested in based on the historical check-in data generated by the user,and then provides a personalized POI recommendation list that the user may visit later from the POIs that the user has not visited.There is a wealth of contextual information available in LBSN,and the geographic information contained in it is particularly important.The geographic information often implies the user’s activity behavior in the real world,which helps to provide users with higher performance POI recommendation.Aiming at the insufficient utilization of geographic factors in POI recommendation,a personal factorization model(Geographic Personal Matrix Factorization,GPMF)that considers geographic information from the perspective of the relationship between users and POIs is proposed.Based on the matrix factorization model,GPMF considers the role of geographic factors from the perspective of the positional relationship among users,the distribution relationship between users and POIs,and the proximity relationship and clustering information among POIs,and mines the influence of geographic factors among different objects and carries out unique modeling.Firstly,for the influence of geographic factors among users,the geographically similar users are calculated according to the user activity center.Then,based on the cosine similarity,the check-in data is used to calculate the behavior similar users.Secondly,for the influence of geographic factors between users and POIs,the center of the user’s activity range is calculated from the user’s check-in data,and then a nonlinear equation is used to determine whether the POI is within the user’s activity range.According to this equation,the influence of geographic factors can be described.Finally,for the influence of geographic factors among POIs,the K nearest neighbors(KNN)algorithm is used to calculate the geographic neighbors of each POI,and the number of visits to the geographic neighbors of the POI is used to calculate the exposure impact.Then,the cluster of POI is divided by density clustering algorithm,and the spatial clustering phenomenon is studied by the cluster of POI and the popularity of POI.By considering the role of geographic information from different aspects and incorporating it into the matrix factorization model,a unified recommendation model is constructed.We conduct comparative experiments with multiple same type recommendation methods on the real-world check-in Foursquare dataset and Gowalla dataset.The results show that the performance of GPMF is better than the current commonly used POI recommendation methods under the Precision and Recall metrics.In addition,through the analysis of the impact of different geographic information modeling methods on the POI recommendation,it is further proved that comprehensive consideration of geographic information from multiple perspectives can help recommendation,and describing geographic information through proximity relationships can improve the performance of recommendation model more effectively than other forms.Therefore,the GPMF model has good performance and validity,and it can alleviate the problem of insufficient use of geographic information.