Estimating NBC-based recommendations on arbitrarily partitioned data with privacy

Abstract

Providing partitioned data-based recommendations has been receiving increasing attention due to mutual advantages. In case of limited data, it is not likely to estimate accurate and reliable predictions. Therefore. e-commerce sites holding insufficient ratings prefer offering predictions to their customers based on integrated data. However, users' preferences about products are considered online vendors' confidential and valuable assets; and they do not want to disclose them their partners during collaborative prediction processes. In order to eliminate privacy, financial, and legal concerns of those companies having inadequate data and want to provide recommendations on combined data, we propose a privacy-preserving scheme to estimate naive Bayesian classifier-based predictions on arbitrarily partitioned data between two parties. Our method helps online vendors provide binary ratings-based predictions on partitioned data without violating their confidentiality requirements. We show that the proposed scheme is secure and able to offer recommendations efficiently. Our real data-based experiments demonstrate that collaboration is vital for better services; and accuracy losses due to privacy measures can be suppressed by the gains due to collaboration. Thus, our method is preferable for estimating accurate predictions efficiently on partitioned data while preserving data holders' privacy over the scheme on split data only