Nonlinear Solution for l_2,1-Norm Based Feature Selection and Neural Network

In joint l_2,1-norm based feature selection methods, predicted attributes shared parallel sparsity coefficients and outliers could be removed. However, previous optimizingl_2,1-norm solutions were usually based on linear solutions which could not discover the non-linear relationship between features. In this paper, we proposed a nonlinear solution for l_2,1-norm based feature selection using the framework of neural network. With combination of l_2,1-norm and neural network, on the one hand, we can make use of the nonlinearity of neural network to optimize l_2,1-norm based problem. On the other hand, the l_2,1-norm can help neural network to achieve feature selection. Finally, the proposed method is compared with famous feature selection methods on eight datasets. Experimental results empirically show the superiority of our method.