Research On The Recommendation Method Of Privacy Protection Points Of Interest

The point-of-interest recommendation system plays an important role in locationbased services.It aims to recommend locations that may be of interest to users on social platforms by analyzing the user's history or combining other information.However,an untrustworthy recommendation system that exposes check-in data is a privacy threat.Attackers may infer some personal information of the user based on the check-in data,such as workplace,religious beliefs,and behavior habits.Therefore,it is of great significance to study a point-of-interest recommendation method that can not only protect user data privacy,but also ensure high usability.The differential privacy technology based on the perturbation idea can provide mathematical proofs to prevent privacy leakage.Applying it to the point of interest recommendation method makes it impossible for the attacker to infer the privacy preference of the user based on the recommendation result,and achieves the purpose of privacy protection.At the same time,the recommendation method combines the auxiliary information of the user and the point of interest(such as social attributes and geographical attributes)to improve the accuracy of recommendations.The main work done in this thesis is as follows:(1)Under the assumption that the data aggregator is credible,based on the centralized differential privacy technology,combined with user preferences and real-time needs,a time-aware privacy protection point of interest recommendation method is proposed.This method includes two algorithms:the privacy protection interest point category prediction algorithm uses matrix decomposition and singular spectrum analysis techniques to discover the time-evolving user category preferences,and noise is added to the gradient during the matrix decomposition process to protect the user's preference data for the category.The privacy protection point-of-interest recommendation algorithm based on weighted HITS weights combined with the influence of social geography to personalize specific points of interest for users,divide the sensitivity according to the check-in frequency of the location,add noise to the user's location check-in score,and protect the user's preference for specific points of interests.The proposed method can achieve a good balance between protecting user sign-in data and recommending performance.(2)Under the assumption that the data aggregator is not credible,based on the localized differential privacy technology,combined with geographic and social attributes,a privacy protection point-of-interest recommendation method combining geographic and social attributes is proposed.The method includes three algorithms:a privacy protection matrix factorization algorithm combined with social networks,a matrix factorization technology based on localized differential privacy,and private learning of implicit user preference features;a geographic correlation calculation algorithm based on kernel density estimation to obtain points of interest Geographical attributes;based on random response mechanism of interest point popularity estimation algorithm to count the visit count of interest points and generate the popularity characteristics of interest points.The results of the weighted synthesis of the three algorithms predict the user's rating of points of interest.While ensuring that users' explicit ratings and implicit preference data are not leaked,a better recommendation effect can be achieved.(3)Privacy analysis and experimental analysis are carried out on the two privacy protection interest point recommendation methods proposed above.Privacy analysis shows that the two.proposed methods meet the definition of differential privacy.Based on the real location data set,a comparative experiment was carried out and two evaluation indicators of precision rate and recall rate were used.The experimental results show that the two methods proposed can achieve better recommendation effects while ensuring the user's preference data.