Cost-Sensitive Encoding for Label Space Dimension Reduction Algorithms on Multi-label Classification

Abstract

Multi-label classification (MLC) extends multi-class classification by tagging each instance as multiple classes simultaneously. Different real-world MLC applications often demand different evaluation criteria (costs), which calls for cost-sensitive MLC (CSMLC) algorithms that can easily take the criterion of interest into account. Nevertheless, existing CSMLC algorithms generally suffer from high computational complexity. In this work, we study a family of MLC algorithms, called label-space dimension reduction (LSDR), which is known to be efficient for MLC but not cost-sensitive. We propose a general framework that directs LSDR algorithms to embed the cost information instead of the label information. The framework makes existing LSDR algorithms cost-sensitive while keeping their efficiency. Extensive experiments justify that the proposed framework is superior to both existing LSDR algorithms and CSMLC algorithms across different evaluation criteria.