Dependence maximization based label space dimension reduction for multi-label classification

High dimensionality of label space poses crucial challenge to efficient multi-label classification. Therefore, it is needed to reduce the dimensionality of label space. In this paper, we propose a new algorithm, called dependence maximization based label space reduction (DMLR), which maximizes the dependence between feature vectors and code vectors via Hilbert–Schmidt independence criterion while minimizing the encoding loss of labels. Two different kinds of instance kernel are discussed. The global kernel for DMLRG and the local kernel for DMLRL take global information and locality information into consideration respectively. Experimental results over six categorization problems validate the superiority of the proposed algorithm to state-of-art label space dimension reduction methods in improving performance at the cost of a very short time.