Learning Common and Label-Specific Features for Label Distribution Learning  with Correlation Information

Abstract

Label distribution learning quantifies the label space for each instance and has wide applicability in different fields. However, most existing work  primarily focuses on common features, ignores the significance of label-specific features. Meanwhile, label correlation is well used in these work, but they still have limitations in capturing instance correlation. Therefore, in this paper, we propose a novel approach for label distribution learning to learn common and label-specific features with correlation information i.e. instance correlation and label correlation. First, we conduct $\ell_{2,1}-$norm regularization to select common and label-specific features, simultaneously. Then, we implement an optimization function to identify instance correlation in feature space. In addition, label correlation is collected with Pearson correlation coefficient.  Finally, we design an objective function that enables the joint learning of common and label-specific features while leveraging correlation information. Comprehensive experiments manifest the superiority of our proposed approach against other well-established label distribution learning algorithms.

Keywords: Label distribution learning, Label-specific features, Common features, Instance correlation, Label correlation