Research On POI Recommendation Algorithm Based On Community Clustering

In recent years,Web2.0 technology has been rapid development,social networks have gradually flourished,such as WeChat,Facebook,Twitter and other applications have been popular in the world.At the same time,with the development of smart phones,support for GPS function and can run the provision of location-based services APP mobile phone began to spread,in this case,location-based social networks(LBSN)will flourish.Based on point-of-interest(POI)recommendation system,a user can find interested businesses and locations,get rich and colorful experience,on the other hand,businesses can make advertising push and commodity marketing,increase turnover and profit.At present,the recommendation algorithm has been widely used in the traditional electronic business platform and social networking platform,for the location of social networks,the recommendation system is also very important.The POI recommendation system contains heterogeneous data,such as friends relationship data,geographic data and POI's comment text and score,the use of these data can effectively improve the accuracy of recommendation system.But LBSN data often have sparse data,and even some information missing,for example,some users did not add friends,so the data sparseness and robustness is the problem that POI recommendation system must be solved;in addition,due to the characteristics of LBSN data is heterogeneous,multi-dimensional,therefore how to fuse these information is also worth to study in the recommended model.In this thesis,we focus on the sparseness of data,the real-time of recommendation,the multi-source heterogeneity of data and the robustness of the algorithm model these problems which POI recommendation system must be solved,and by modeling the user's access to the POI behavior to predict the user's preferences,so as to recommend POI for user.Specifically,the research work and achievements of this thesis are mainly embodied in the following:(1)In order to solve the problem of data sparsity in the POI recommendation system,a pre-filling algorithm PACC based on Compatible Class is given in this thesis.In this algorithm,the compatibility relation is used to replace the non discernible relation,and the sparsity of the user-poi-rating matrix is reduced.At the same time,the time complexity of the algorithm is analyzed and comparative experiments were carried out on the Yelp data set.(2)In order to overcome the real time problem and the defects of the traditional community discovery algorithm applied to the recommendation system,an interest preference modeling algorithm based on community clustering algorithm CDCF is given.CDCF algorithm combines the user's preference information and social friends relation,and by introducing community clustering method to model in advance,and to reduce the neighbor search space,improve the accuracy of recommendation system and real-time.(3)The POI recommendation system contains rich heterogeneous data,which can effectively improve the performance of the recommendation system.In this thesis,by adding the distance factor for CDCF community clustering algorithm,we give a joint model SoGeoSco(Social,Geographical and Score)to predict the ratings by fusing the data of social relationships,geographic location data and user-poi-rating matrix.In the SoGeoSco model,the user's access probability is determined by three factors: the distance between the user and the POI,the individual interest and the social interest.Specifically,SoGeoSco model gets the distance between user and POI on geography data by introducing Naive Bayesian classifier,using the CDPC algorithm respectively to social relationship data and sign rating data modeling to get personal interest and social interest,finally,a robust rule is used to fuse the multi-source data to get the SoGeoSco model.A contrast experiment was performed based on publicly available data sets,the result of analysis shows that,compared with other mainstream interest recommendation algorithm,SoGeoSco model can improve the accuracy and recall rate,and still has a good performance when some information is missing,showing the robustness,and obtain better recommendation result.