Abstract: Label distribution learning(LDL)is an innovative machine learning paradigm that focuses on the label descriptiveness of specific instances. Existing LDL models were built on a simple feature space where all features were shared by all class labels. However, this traditional data representation strategy tended to select features that were distinguishable across all labels while overlooking label-specific features that were relevant and distinctive for each class label. In this paper, we proposed a label distribution learning method based on label-specific features and common features. First, we calculated the label correlation matrix for label pairs, and for each label pair, all instances were partitioned based on descriptiveness, clustering was performed for each class, and label pair-specific features were constructed. Next, clustering analysis was performed on all instances to construct common features for the labels. Finally, training and testing were conducted by querying the clustering results. Extensive experiments on multiple real label distribution datasets validated the superiority of the method and the effectiveness of label-specific features.

Key words: label distribution learning, label-specific feature, label common features, label correlation, clustering