Feature selection based on correlation deflation

Si-Bao Chen, Chris H. Q. Ding, Zhili Zhou, Bin Luo

Abstract: Feature selection is very important in many machine learning and data mining applications. In this paper, a simple and effective correlation-deflation-based feature selection method is proposed. The objective function of residual minimization constrained by $$L_{2,0}$$ L 2 , 0 -norm is proved to be equivalent to maximizing sum of square of correlations between class labels and features. Then the whole procedure of correlation-deflation-based feature selection turns into selecting features out one-by-one by deflating correlations. Experiments on several public benchmark data sets show that the proposed method has better residual reduction and classification performance than many state-of-the-art feature selection methods.