Proposal of Collaborative Filtering Focusing on Feedback on Specific Items

Abstract

This paper proposes a method for matrix factorization-based collaborative filtering (CF) that focuses on feedback to specific items. Although a large amount of user information is usually required for accurate recommendation, privacy concern makes it difficult to collect user information. As a solution, we focus on applying CF to a small amount of feedback on a few items by developing a method for determining the items for which user feedback can be efficiently obtained (Probing Items), and a method for learning a recommendation system focusing on the probing items. This paper focuses on the latter and proposes a method for learning a matrix factorization-based CF, in which probing items are considered at every epoch while other items are randomly dropped. Experimental results show the effectiveness of the proposed method.