Similar or Dissimilar Users? Or Both?

Abstract

E-commerce sites utilize collaborative filtering (CF) techniques to offer recommendations to their customers. To recruit new customers and keep the current ones, it is imperative for online vendors to provide accurate predictions efficiently without deeply violating users' privacy. To improve the overall performance of CF systems, it is important to use the appropriate data. We investigate how to improve naive Bayesian classifier (NBC)-based CF systems' online performance. For this purpose, we group users in various clusters so that predictions can be generated on similar or dissimilar; or both groups of users' data. Grouping users into clusters makes it possible to utilize smaller amount of data. We perform real data-based experiments to assess how overall performance changes with different data. Our results show that online time to generate referrals improves significantly when clustering is utilized to get proper data.