Learning Common and Label-Specific Features for Multi-Label Classification With Missing Labels

Abstract

Multi-label learning is a subfield of machine learning that addresses the issue of each instance belonging to numerous class labels at the same time. However, in some real applications, we can only receive a partial set of labels for each instance due to the difficulty and high cost of labeling data. The vast majority of existing multi-label classification methods on missing labels rely on first- or second-order label correlation learning to fill in the original label space while building multi-label learning models with label-specific features; nevertheless, the single label correlation learning mechanism used in these methods is insufficient to maintain the consistency of the feature-label space. To address this issue, we propose the CLSML approach, which incorporates higher-order label correlation learning constraints in the classifier training model to complete missing labels while training the classifier. In addition, to improve the consistency of the feature-label space, we develop a two-stage second-order label correlation learning technique based on cosine similarity to further confine the label output. Furthermore, we employ the $l_{1}$ -norm regularizer to learn label-specific feature representations, followed by the $l_{2,1}$ -norm regularizer to constrain the row sparsity of the classification matrix and select label-common features. Experimental results comparing ten cutting-edge multi-label learning algorithms with missing labels on fourteen multi-label benchmark datasets demonstrate the effectiveness of our suggested approach.