A Normative Approach to Privacy-Preserving Recommender Systems: Integrating Matrix Factorization and Genetic Algorithms

Abstract

As recommendation systems heavily depend on user data, these systems are susceptible to potential privacy breaches. To mitigate this issue, differential privacy (DP) protection techniques have been developed to offer robust privacy safeguards. Nevertheless, a majority of the extant DP-based recommendation algorithms tend to introduce excessive noise, consequently impairing the quality of recommendations. In response, this study presents a novel DP-preserving recommendation algorithm that integrates matrix factorization (MF) and a genetic algorithm (GA). Initially, the MF problem is transformed into two interrelated optimization problems, namely, the user-hidden factor and the item-hidden factor. Subsequently, GA is employed to address these optimization issues. An enhancement index mechanism is incorporated into the individual selection of GA, while the variation process of GA is devised based on identifying crucial hidden factors. Utilizing the enhancement index mechanism aids in minimizing the algorithm’s perturbation level, thereby achieving an optimal balance between privacy protection and algorithm utility. Experimental analyses, encompassing recommendation accuracy, efficiency, and parameter variation effects, are conducted on Last.fm and Flixster datasets. The findings corroborate that the proposed system outperforms existing alternatives under stringent privacy constraints, thereby attesting to its efficacy.