Confident Collaborative Metric Learning

Abstract

Modern recommendation systems use embedding for secondary applications and implicit feedback data for learning. In recent years, collaborative metric learning (CML), a method that can precisely capture the relationship between users and items, has been developed for the first requirement. However, CML with implicit feedback data suffers from noisy-label issues. This is mainly because CML and the related works only consider few types of noise. To overcome these limitations, we develop a method for estimating the noise rate and excluding extremely noisy user–item pairs from the data prior to learning CML. Experimental results show that the proposed method significantly outperforms existing methods in terms of ranking metrics.