A Comparative Study on Single and Dual Space Reduction in Multi-label Classification

Abstract

Multi-label classification has been applied to several applications since it can assign multiple class labels to an object. However, its effectiveness might be sacrificed due to high-dimensionality problem in both feature space and label space. To address these issues, several dimensionality reduction methods have been proposed to transform the high-dimensional spaces to low-dimensional spaces. This paper aims to provide a comprehensive review on ten-dimensionality reduction methods that applied to multi-label classification. These methods can be categorized into two main approaches: single space reduction and dual space reduction. While the former approach aims to reduce the complexity in either feature space or label space, the latter approach transforms both feature and label spaces into two subspaces. Moreover, a comparative study on single space reduction and dual space reduction approaches with five real-world datasets are also reported. The experimental results indicated that dual space reduction approach tends to give better performance comparing to the single reduction approach. Furthermore, experiments have been conducted to investigate the effect of dataset characteristics on classification performance.