Semi-supervised Multi-label Learning Algorithm Using Dependency Among Labels

Abstract

In this paper, we present a semi-supervised algorithm for multi-label learning by exploring the relationship among labels. Based on the accuracy, we determine the classification order for labels, a list of classifiers is trained by this order, with each classifier being trained by using the outputs of the previous classifiers in the list as additional input features. Experiments on three multi-label data sets show that our algorithm has substantial advantage over the comparing algorithms.