Self-expression multi-label feature selection based on fuzzy decision

Abstract

The large amount of high-dimensional data poses a great challenge to multi-label learning. Feature selection is an effective method to alleviate this problem. However, many existing multi-label feature selection models either ignore the intrinsic spatial structure of samples or have no restrictions on the predicted label values. To solve the above problems, a sample self-representation multi-label feature selection method based on fuzzy decision is proposed in this paper. Firstly, a self-representation coefficient matrix of samples is proposed, which not only retains the original data structure information, but also reflects the distribution structure of data. Then, a fuzzy decision function is introduced to fuzzy prediction labels which well represents the membership of a sample to a class and is more consistent with the real label distribution. The $L_{2,1}$-norm is imposed on the feature weight matrix to ensure sparsity and the $F$-norm is introduced into the self-expression matrix to weaken the effects of redundancy and anomalous samples. Finally, the gradient descent method is used to optimize the objective function. Experimental results on 12 multi-label datasets show that the proposed method performs better than other state-of-the-art multi-label feature selection methods, and obtain a significant increase in classification accuracy of about 2%–3% over all the compared approaches.