Missing multi-label learning with non-equilibrium based on classification margin

Abstract

Multi-labels are more suitable for the ambiguity of the real world. However, missing labels are common in multi-label learning datasets; this results in unbalanced labeling and label diversity, which directly affect the performance of multi-label learning. Therefore, the classification and modeling of imbalanced data in missing multi-label learning are problems that need to be urgently solved Current methods mostly focus on combining sampling techniques with cost-sensitive learning and incorporating label correlation to improve the performance of the classifier, but generally they do not consider label loss caused by label cost. In fact, labeling unknown instances is often affected by the threshold of the discriminant function, especially for the label types near the threshold. Based on our previous research, we believe that information such as data distribution density and label density can be integrated into the label correlation, and that the classification margin can be expanded to effectively solve the labeling quality of labels near the threshold. Therefore, in this paper we propose a non-equilibrium multi-label learning algorithm based on the classification margin and aimed at completing the missing labels. First, the classification margin is proposed, and the label space is expanded by the label density. Then, the information entropy is used to measure the correlation between labels, and the label confidence matrix is constructed. The label confidence matrix is then unbalanced using the positive and negative label density, and the non-equilibrium label confidence matrix is used for label completion to obtain an informative label completion matrix. Finally, the kernel extreme learning machine and the label completion matrix are used for linear prediction. The experimental results show that the proposed algorithm has some advantages over other multi-label learning algorithms.