An approach for multi-label classification by directed acyclic graph with label correlation maximization

Abstract

Traditional supervised learning approaches primarily work in the single-label environment. However, in many real-world problems, data instances are usually associated with multiple labels simultaneously, and multi-label learning is increasingly required in many modern applications. In multi-label learning, the key to successful classification is effectively exploiting the complex correlations among the output labels. This paper proposes a novel multi-label learning method inspired by the classifier chain approach. The main contribution of this work is to model the correlations of the labels using a directed acyclic graph. Starting from the simple intuitive notion of measuring the correlations among the labels, the proposed method is designed as a multi-label learning method that maximizes the correlations among labels. To evaluate its effectiveness, the proposed method is compared with the state-of-the-art approaches. Extensive experiments demonstrated the proposed method to be highly competitive with the other multi-label approaches.