Collaborative filtering algorithm based on differential privacy technology

Abstract

The invention provides a collaborative filtering algorithm based on the differential privacy technology. The algorithm comprises the following steps: scrambling aiming at a target function of an optimization algorithm ALS, and providing a mathematical derivation procedure of the parameter solution of the ALS target function scrambling algorithm meeting the differential privacy at the moment; the noise quantity added in a recommendation system is greatly lowered, the availability of a recommendation result can be guaranteed while guaranteeing that the recommendation system meets the differential privacy.

Description

technical field [0001] The present invention relates to the field of item recommendation algorithms, and more specifically, to a collaborative filtering algorithm based on differential privacy technology. Background technique [0002] The recommendation system relies on the user's personal information, so there is a risk of exposing personal privacy when using the recommendation system. In recent years, several research works have been devoted to enhancing the privacy protection of recommender systems. Canny proposes a decentralized storage of user information and divides it into multiple communities, so attackers must target multiple systems to attack. Polat et al. proposed that by adding uncertainty to user ratings through random scrambling, even if an attacker wants to attack user data, only part of the data will be leaked. Nikolaedko et al. proposed to apply multi-party computing to matrix factorization, so that recommendation learning only uses item information withou...

AI Technical Summary

Problems solved by technology

The second is that McSherry et al. used differential privacy in the collaborative filtering recommendation system. They performed differential privacy processing on the item-to-item covariance matrix, and divided the recommendation system into a learning phase and a prediction phase (that is, recommendation). Implementing differential privacy protections is feasible without severe loss of recommendation accuracy, but they do not model latent factors

[0004] The main shortcomings of the existing research are: (1) Most of the recommendation systems based on differential privacy add a large amount of noise, which has a great impact on the usability of the final recommendation results; (2) Although some researchers put differential privacy Conservation techniques are applied to latent factorization of matrix factorization, but a rigorous mathematical proof is lacking

Images