Crorank: Cross Domain Personalized Transfer Ranking For Collaborative Filtering

Collaborative filtering techniques aim at recommending products to users based on their historical feedback. And many algorithms focus on personalized ranking problem with implicit feedback due to the "one-class" nature of many real-world datasets in a variety of services. Most of the existing personalized ranking methods are confined to one domain of data source and the question of how to model users' preferences information across distinct domains is usually be ignored. There are some transfer learning approaches that try to transfer numerical ratings, auxiliary social relations and other information across different domains but they do not address how users' preferences information varies from one domain to another accordingly. And they mainly exploit rating prediction problem rather than personalized ranking problem. In this paper, we propose an algorithm called CroRank to address the question, "How to bridge users' preferences information across different domains to promote better personalized ranking performance?". There are two main steps in CroRank, we first present an algorithm called multiple binomial matrix factorization (MBMF) to bridge the gap between items from distinct sources and then we introduce transfer Bayesian personalized ranking (TBPR) to recommend items for each user in the target domain. In CroRank, users' inclinations can transfer from the auxiliary domain to the target domain to provide better personalized ranking results. We compare CroRank to the state-of-the-art non-transfer models to demonstrate the improvements in flexibility and effectiveness.