I. Introduction
With rich information presented in multimedia data, an object can be naturally associated with multiple labels. Therefore, many real-world classification tasks are not limited to multi-class classification (i.e., only one category label for each input instance), but focus on multi-label classification, that is, assigning multiple labels to each instance at the same time. Multi-label learning can be applied to many fields, such as medical diagnosis recognition [1], [2], natural image recognition [3], [4], facial action units recognition [5]–[7] and human attribute recognition [8], [9], etc. In addition to the application of image modality, increasing attention has been paid to non-image modality, especially for multi-label semantic decoding in non-image modality.