Privacy-preserving Eigentaste-based collaborative filtering

Abstract

With the evolution of e-commerce, privacy is becoming a major concern. Many e-companies employ collaborative filtering (CF) techniques to increase their sales by providing truthful recommendations to customers. Many algorithms have been employed for CF purposes, and Eigentaste-based algorithm is one of them. Customers' preferences about products they purchased previously or showed interest are needed to provide recommendations. However, due to privacy concerns, customers refuse to contribute their ratings at all; or they might decide to give false data. Providing truthful referrals based on such inadequate and false data is impossible. Therefore, providing privacy measures is vital for collecting truthful data and producing recommendations. In this paper, we investigate how to achieve CF tasks (predictions and top-N recommendations) using Eigentaste, which is a constant time CF algorithm, without greatly exposing users' privacy. To accomplish privacy, we employ randomized perturbation techniques (RPT). We modify and/or simplify original Eigentaste algorithm in such a way to provide private referrals efficiently with decent accuracy. We investigate our proposed schemes in terms of privacy. To evaluate the overall performance of our schemes, we conduct experiments using real data sets. We then analyze our outcomes and finally provide some suggestions.