Class label fusion guided correlation learning for incomplete multi-label classification

Abstract

Label correlation learning is a challenging issue in multi-label classification, which has been extensively studied recently. Typically, the second-order label correlation is achieved by fusing information from pairwise labels, while high-order correlation arises from integrating global information of the entire label matrix with the help of some regularization constraints. However, few studies focus on collaboratively learning label correlations through local and global label fusion. Unfortunately, in the case of label missing, neither second-order nor high-order label correlations can be accurately measured and characterized. To address the two issues, a novel approach for incomplete multi-label classification called class label fusion guided correlation learning (CLFCL) is proposed. The pointwise fuzzy mutual information is introduced for prior fusion of paired labels. Specifically, the second-order label correlation is obtained by relaxing the pointwise mutual information. Simultaneously, an adaptively low-rank regularization technique is developed to fuse globally relevant labels so as to extract the high-order correlations. By integrating second-order and high-order label correlations, the label distribution of instances is learned. To recover missing labels, a multi-label classifier is trained by regressing features to label distribution space rather than original logical label space. An efficient algorithm is designed to solve the built nonconvex optimization. Extensive experimental results validate the superior performance of the proposed model against state-of-the-art missing multi-label classification methods.