1 Introduction
Label ambiguity [1] is the phenomenon that one instance is related to multiple labels by different degrees, which is a hot topic in the field of machine learning. Take multi-label image classification and emotion recognition as examples. Fig. 1a shows a multi-label scene image from [2]. Note that “Water” has higher importance than “Sun”, although both are positive labels. Fig. 1b shows an image from the JAFFE database [3] with a ground-truth single-label “ANG.”. However, the image is a mixture of many kinds of emotions by different relevance. Traditional supervised learning paradigms, such as Single-Label Learning (SLL) and Multi-Label Learning (MLL) [4], model the correspondence between instances and labels by 0 or 1, which fails to consider label ambiguity.