I. Introduction
Learning with ambiguity has gained increasing attention in recent years, becoming a popular topic in machine learning. In multilabel learning (MLL), an instance is associated with multiple labels, which partially addresses the label ambiguity problem [1], [2], [3]. MLL assigns logical labels (0 or 1) to each instance, separating the candidate labels into relevant and irrelevant. However, the importance degrees of relevant labels for an instance may differ, and thereby, using the same logical label (i.e., 1) for the relevant labels may not be suitable for real-world tasks. For example, Fig. 1 shows two posters from the movies of “The Great Gatsby” and “The Godfather” (taken from the Twitter-LDL dataset [4]). Below each poster is the averaged emotion distribution annotated by different annotators, where the -axis represents the eight different emotions and the -axis shows the importance degree of each emotion. In Fig. 1(a), both “Amusement” and “Contentment” are positive labels, but their importance degrees are different. Similarly, in Fig. 1(b), “Anger” and “Sadness” are positive labels, but their importance degrees also vary. Thus, assigning logical labels to an instance in MLL ignores the relative importance of labels.
Example of the relative importance degrees between relevant labels of different instances. (a) and (b) are two posters from the movies “The Great Gatsby” and “The Godfather,” respectively, which are from the Twitter-LDL dataset [4].