Semi-supervised imbalanced multi-label classification with label propagation

Multi-label learning tasks usually encounter the problem of the class-imbalance, where samples and their corresponding labels are non-uniformly distributed over multi-label data space. It has attracted increasing attention during the past decade, however, there is a lack of methods capable of handling the imbalanced problem in a semi-supervised setting. This study proposes a label propagation technique to settle the semi-supervised imbalanced multi-label issue. Specially, we first utilize a collaborative manner to exploit the correlations from labels and instances, and learn a label regularization matrix to overcome the imbalanced problem in the labeled instance. After that, we extend to semi-supervised learning and explore to represent the similarity of instances with weighted graphs on labeled and unlabeled data. Then, the data distribution information and label correlations are fully utilized to design the loss function under the consistency assumption manner. At last, we present an iterative scheme to settle the optimization issue, thereby achieving label propagation to address the imbalanced challenge. Experiments on a variety of multi-label data sets show the favorable performance of the proposed method against related comparing approaches. Notably, the proposed method is also validated to be robust with a limited number of training instances.