Privacy-Preserving Svd-Based Collaborative Filtering on Partitioned Data

Abstract

Collaborative filtering (CF) systems are widely employed by many e-commerce sites for providing recommendations to their customers. To recruit new customers, retain the current ones, and gain competitive edge over competing companies, online vendors need to offer accurate predictions efficiently. Therefore, providing precise recommendations efficiently to many users in real time is imperative. Singular value decomposition (SVD) is applied to CF to achieve such goal. SVD-based CF systems offer reliable and accurate predictions when they own large enough data. Data collected for CF purposes, however, might be split between different companies, even competing ones. Some vendors, especially newly established ones, might have problems with available data. To increase mutual advantages, provide richer CF services, and overcome problems caused by inadequate data, companies want to integrate their data. However, due to privacy, legal, and financial reasons, they do not want to combine their data. In this article, we investigate how to provide SVD-based referrals on partitioned (horizontally or vertically) data without greatly jeopardizing data holders' privacy. We conduct real data-based experiments to assess our schemes' overall performance and analyze them in terms of privacy and supplementary costs. Our results show that it is possible to provide accurate SVD-based referrals on integrated data while preserving e-companies' privacy.