Transfer Learning for Heterogeneous One-Class Collaborative Filtering

Abstract

Various memory- and model-based collaborative filtering algorithms have been designed for multiclass feedback (such as grade scores) in the past two decades. Recently, one-class feedback (such as positive feedback and implicit examination) has been recognized as a more pervasive and important source of information in many real recommendation systems. Previous work along these lines mainly focus on homogenous one-class positive feedback, such as likes on Facebook or transactions on Amazon, which might not capture a user's true preferences due to the sparsity of such data. To alleviate this sparsity problem, the authors study positive feedback and implicit examinations simultaneously, coined as heterogeneous one-class collaborative filtering (HOCCF). Specifically, they designed a novel transfer learning algorithm for HOCCF, called transfer via joint similarity learning (TJSL), that jointly learns a similarity between a candidate item and a preferred item, and a similarity between a candidate item and an identified likely-to-prefer examined item. Joint similarity learning has the merit of being able to connect two seemingly unrelated items along sparse positive feedback only. Empirical studies on three real-world datasets show that TJSL can recommend items more accurately than other state-of-the-art methods.