Multi-Label Learning Based on Clustering Optimization

Abstract

Multi-label learning has been widely used in various fields. One problem in multi-label learning is that mining the correlation information between class labels in the entire label space is not only computationally unaffordable, but may be inappropriate (possibly inconsistent with the real world), which leads to the incorrect introduction of some label dependencies. So, the common solution is to divide the whole label space into several subspaces. However, the process of sub-space partitioning is usually independent of the classifier training process in the existing multi-label learning algorithms, which leads to the division is not optimal for the next classifiers training. In order to solve the above-mentioned problems, we propose a method whose basic idea is to optimize the label clustering process and the classifiers training process simultaneously. In addition, we apply group lasso constraints to the coefficient matrix to ensure that the output of classifiers can maintain the characteristics of consistency within each label cluster and mutually exclusive between label clusters. Experiments conducted on five benchmark multi-label datasets demonstrate the competitive performance of our proposed method compared with several state-of-the-art methods.