I. Introduction
Traditional supervised learning paradigms, such as single-label learning (SLL) and multilabel learning (MLL), assume that labels are certainly relevant or irrelevant to instances. However, in many real-world problems, the relation between instances and labels often contains somewhat uncertainty [1], known as label ambiguity [2] in the field of machine learning. For instance, Fig. 1 shows an example image from the JAFFE [3] database with the ground-truth label “ANG.” The image blends multiple emotions with different relevance, which include not only “ANG” but also some other emotions with less relevance. However, SLL and MLL use 0 and 1 to represent the relation between instances and labels, which ignores label ambiguity.
Example image from the JAFFE [3] database with the ground-truth label “ANG.” The mean ratings are normalized into label distribution.