Research On Cross-Domain Fusion And Privacy-Preserving In POI Recommendation

With the promotion of the Internet and the rapid growth of data collection,information overload makes it difficult for users to search for information and directly reduces user experience.Therefore,the research on recommender system has attracted extensive attention.Point-of-interest(POI)recommendation based on location is different from the traditional product recommendation because it has strict requirements on location and time.The limited check-in ability of users makes the location data extremely sparse compared with the traditional product recommendation,and it often needs to analyze the context information of users to realize effective recommendation.In addition,the cold-start problem is also an important factor that reduces the recommendation efficiency.At the same time,due to privacy concerns,users are often reluctant to share their check-in data,which also makes it difficult for POI recommendation platform to collect user's check-in data.In order to solve the above problem,POI recommendation mostly fused user's comment text,time and location factors to relieve the POI data sparseness.Most of the cross-domain fusion methods based on multi-dimensional data control their influence by setting a linear weight.And there are few cross-domain fusion methods to consider the privacy protection of users.Therefore,this paper carried out a study on the privacy-preserving issues of cross-domain fusion in POI recommendation.Integrated federated learning framework proposed a cross-domain POI recommendation algorithm for privacy protection,which can effectively integrate multi-dimensional source data and protect users' privacy security.The main research work is as follows.(1)We first propose a cross-domain POI recommendation method based on federated learning and privacy-preserving.On the one hand,we integrate data from the auxiliary domain into user's interest analysis to improve the cold-start problem.And take the data of the electronic commerce field as an example to analysis.On the other hand,for privacy issues in data fusion,we use the federated learning framework and store user's data locally.What's more,we use the encrypted feature distribution into knowledge transfer to realize the privacy-preserving.In this method,we keep all user data locally on the client and only interact the encrypted feature distribution.On the server side,an improved neural network is used to train the latent features of the encrypted feature distribution to realize feature alignment between two domains.Finally,the model comparison proves that the method can effectively improve the recommendation accuracy while protecting the user's privacy.(2)Secondly,we further consider the scenario of multi-source data source integration for the complex association between cross-domain for the complex relationship between cross-domain.Integrating the information of users in e-commerce field and social network dimension,and introduce the attention mechanism to pay attention to the importance of the different relation information,so that the server can learn user preferences adaptively.And form a cross-domain intelligent POI recommendation algorithm with multi-source data,while ensuring user's privacy security.(3)Then,this paper designs a corresponding cross-domain fusion and privacy protection algorithm,and theoretically analyzes the performance of the algorithm from privacy security and algorithm complexity.And prove that the proposed method can effectively protect the privacy security in cross-domain POI recommendation.(4)Finally,this paper first conducts experimental verification based on the public dataset Movielens and Foursquare.And evaluates the accuracy of recommendation results from Precision and Recall by comparing the existing methods.Then validate the cross-domain recommendation methods fused the social network by Epinions and Four Square.Experimental results show that the proposed method can effectively improve the recommendation accuracy of POI recommendation in the case of sparse data,and also reduce the running time of the model to some extent.