Research On Privacy Protection Methods In Data Fusion Recommendation

With the rapid development of the Internet,the information increases exponentially on the Internet,and the difficulty of filtering the content they are interested in from the massive and complex data is also increases.Recommendation system has been widely used as an effective information filtering tool.Its goal is to accurately predict the user’s preference for items or information,so as to prioritize content that is more valuable to users and help them make quick decision.Collaborative filtering is widely used in recommendation algorithms.It makes recommendations based on the assumption that similar users have similar preferences.However,when users have little or no historical rating data,the recommendation system will face the problems of data sparsity and cold start.Therefore,more and more researchers propose to integrate multi-source data into the recommendation system as auxiliary information to improve the performance of the recommendation.However,the historical feedback data collected by the recommendation system,such as user ratings,inevitably involves users’ privacy information.With the improvement of users’ awareness of privacy protection,people are paying more and more attention to whether their privacy can be protected while enjoying recommendation services.However,in recommendation based on data fusion,the privacy problem is more challenging because it involves multi-source data and their correlations brings stronger background knowledge to the attacker.Therefore,this paper analyzes the privacy problems in the data fusion recommendation and the shortcomings of existing methods,and proposes the corresponding privacy protection solutions for the data fusion recommendation.The main research results are as follows:(1)Aiming at the problem that the privacy protection method in the news recommendation system integrating knowledge graph cannot balance the noise addition and recommendation effect,we propose a two-stage privacy protection mechanism for a news recommendation with knowledge graph and privacy protection,which is designed based on the differential privacy model.Firstly,privacy budget is dynamically allocated according to the influence weight of feature vectors on recommendation results,and Laplacian noise is added to user feature vectors to achieve privacy protection and reduce the data loss caused by excessive noise addition,so as to improve the utility of data.Secondly,unified Laplace noise is added to the weighted user feature vector to ensure the security of user data.Finally,experiments on real news data sets show that the proposed method can ensure the prediction performance of the model while protecting users’ privacy.(2)Aiming at the data sparsity of recommendation system integrating sentiment analysis,we propose an improved opinion mining model based on neural network,which can integrate the inherent attribute characteristics hidden in the user’s ID representation to enhance the recommendation performance by reducing data sparsity.In addition,we propose a differential privacy neural-network based Opinion mining model,DPNeu O to solve the privacy problems in the recommendation scenarios with sentiment analysis.In this model,noise is added to protect privacy during the process of obtaining optimal parameters from the minimized loss function of gradient descent based on differential privacy model.Finally,a comparative experiment is designed on two real data sets(Amazon and Yelp)to verify the proposed method,which shows that the proposed method can improve the accuracy of recommendation while protecting privacy and security compared with the existing work.