A New Kernel-Based Classification Algorithm for Multi-label Datasets

Abstract

With the emergence of rich online content, efficient information retrieval systems are required. Application content includes rich text, speech, still images and videos. This content, either stored or queried, can be assigned to many classes or labels at the same time. This calls for the use of multi-label classification techniques. In this paper, a new kernel-based multi-label classification algorithm is proposed. This new classification scheme combines the concepts of class collaborative representation and margin maximization. In multi-label datasets, information content is represented using the collaboration between the existing classes (or labels). Discriminative content representation is achieved by maximizing the inter-class margins. Using public-domain multi-label datasets, the proposed classification solution outperforms its existing counterparts in terms of higher classification accuracy and lower Hamming loss. The attained results confirm the positive effects of discriminative content characterization using class collaboration representation and inter-class margin maximization on the multi-label classification performance.