PC-GEO:Potential Check-ins And Geographical Model Based Point-of-interest Recommendation

As a technology that can effectively reduce the overload information on the Internet,personalized recommendation technology is widely used in new media platform,such as e-commerce and social networking for its efficient and accurate displaying in multiple fields(music,movies,books,commodities)..In recent years,with the open of positioning system GPS,positioning technology more and more civilian,location-based social networks(LBSNs)was born and developed.LBSNs attracted a lot of users with its innovative and interesting location services.With the large-scale influx of users,the website collects massive amounts of user information,such as social information,location of interest points,and topic information,which provide important support for mining online behaviors of users.In this context,in order to provide users with more accurate location services,the personalized recommendation technology is introduced into the LBSN.So,point-of-interest(POI)recommendation based on location social networks has become a recommendation field.Important emerging research direction.In LBSN,the user-interest point check-in frequency matrix is rather sparse,and the POI recommendation faces more serious data cold start than the traditional recommendation.In order to alleviate the impact of this problem on the efficiency of the recommendation,researchers began to study various algorithms to improve the efficiency of the recommendation.Among them,one effective mechanism is to combine one or more factors in context information such as geographical influence,social information,and time factor.To improve the performance of POI recommendation,this paper that focuses on the problem of cold start caused by sparse user data proposes a framework PC-Geo(Potential Check-ins and Geographical Model based Point-of-Interest Recommendation)for integrating geographic influences,social information,and time factors.There are four main works of in our paper,as follows:(1)The users' check-in behavior in LBSN is affected easily by related users.The points of interest that related users check in are more likely to be accessed by users.Based on this phenomenon,this paper designed a potential interest learning algorithm to learn the potential interests of users.The algorithm learns a set of POIs from the check-ins of related users(geo-related users,social-related users,check-in similar-related users)as potential interest points of the target user,and then put user's potential POIs fill in the original user-poi matrix,which makes the matrix from only contains two types of data(check-in information and unchecked information)into contains three types of data(check-in information,potential check-in information,and unchecked in information,alleviating matrix sparseness.(2)In the potential POIs learning algorithm: when looking for a similar behavior user,we analyzed the users' check-in behavior in time series,and the time factor is introduced into the calculation of the user similarity;Among many POIs that the relevant user has checked in,we use a maximum value selection strategy and meta-path selection strategy to select some POIs as potential interest for the user,and the two strategies are compared.(3)For the user-poi matrix in 1),we learn latent user feature and latent location feature by matrix factorization technique,and then obtain the preference prediction matrix through the matrix dot product.We capture the geography model by the variable kernel density estimation.Finally,we use a line model to merge the matrix of user preference,geography model to complete the recommendation.(4)In order to evaluate PC-Geo model,the paper compares the PC-Geo model with other excellent algorithms;compares different internal selection strategies;compares the results of potential POIs of different values;compares the PC-Geo model with the random POIs selection by using the publicly available Foursquare dataset and Gowalla dataset.After a series of different experimental results,the validity of the proposed PC-Geo model is proved.