Recommendation method based on label and differential privacy protection

Abstract

The invention provides a recommendation method based on label and differential privacy protection. The method aims to introduce a label concept, similarity of the tags through a co-occurrence principle of the tags is calculated. The tag similarity is used for replacing the Euclidean distance of fuzzy c-means clustering to carry out fuzzy c-means clustering on the tags, so that the tags belong to different clusters. The problem of low recommendation accuracy caused by the hard clustering problem of conventional clustering is solved, and noise conforming to Laplace distribution is added in the clustering process to achieve the purpose of protecting the privacy of a user. And fuzzy c-means clustering is carried out on the tags, so that the problem of data availability reduction caused by directly adding noise into the tag interest degree vectors is further solved, and the privacy of a user is protected while the recommendation accuracy is improved.

Description

technical field [0001] In the field of network information search and processing, especially a recommendation method based on tags and differential privacy protection. Background technique [0002] With the rapid development of information technology and data mining technology, information is growing explosively, and personalized recommendation algorithms have also been widely developed and applied; among them, the most widely used and the best effect is the collaborative filtering algorithm, which is based on user collaboration. Filtering is based on project-based collaborative filtering. These are calculated based on user rating matrices, but they still face the problem of data sparsity. When the rating data is sparse, it is difficult to get similar users, and the quality of recommendation will also decline. The main reason for the above problems is that the amount of data is not enough, and more suitable and easy-to-obtain data information is needed for calculation. An i...

AI Technical Summary

Problems solved by technology

However, since the personalized recommendation system collects a large amount of user information to mine the interest model, this causes the leakage of user privacy. Therefore, preventing user privacy leakage is an urgent problem to be solved in the recommendation system and has become a hot research topic.

[0003] At present, the privacy protection methods of recommendation systems are generally divided into three categories: data perturbation, data encryption, and data generalization; although the perturbation method is relatively simple, it has the problem of weak protection ability; data encryption technology is especially homomorphic encryption. There are many technologies, multi-party computing in terms of security, so that it can be used in collaborative filtering, but it also has the problem of a large size after the encryption algorithm is complex and the public key is generated, and there is no strict definition of privacy. The data generalization algorithm does not The background knowledge mastered by the attacker is defined, and it needs to be continuously improved when encountering new attacks; compared with these traditional privacy protection methods, there are two problems: the inability to strictly prove the level of privacy protection and the security related to the background knowledge mastered by the attacker. The first major flaw is to propose a strictly provable ε-differential privacy protection model. His strict provability is based on strict definitions and solid mathematical foundations. The model assumes that the attacker can master the maximum background knowledge, such as The attacker has obtained all records except the target record, so there is no need to consider the attacker's mastery of background knowledge, and the privacy protection has been quantitatively evaluated, and the budget parameter can be used for comparison

[0004] At present, there are few researches and achievements on the application of differential privacy in the recommendation algorithm. It was first applied to the traditional covariance matrix to predict its score, and later it was applied to matrix decomposition. These are all predictions based on scores. At present, The research and methods in the field of label data are basically blank. Compared with the user privacy of score leaks, the problem of label leakage is more prominent. Therefore, this invention proposes a method for differential privacy protection based on labels. Mining user interests, improving recommendation accuracy while protecting user privacy

Images