Deep unfolded IRLS-ADMM network for classification and sparse feature selection

Two main branches of machine learning methods are model-based methods and deep neural networks. Model-based methods can explicitly include prior knowledge into the model at expense of difficulties in inference, while neural networks are featured in their strong predictive power and straightforward inference approach with the lack of model interpretability. To construct models which are entitled with the advantages of both methods and overcome their problems, the deep unfolding strategy has been developed recently. This paper adopts the idea of deep unfolding to construct a classification and feature selection method. The proposed method is based on the sparse classification; and the iterative inference process of the sparse classification is unfolded into a layer-wise structure analogous to a neural network. Thus, the architecture of our network is fully motivated by the sparse classification method. Different from other neural networks which are essentially black-box methods, our deep unfolded network acts as white-box that features selected in the predictive model can be explicitly returned. Experimental results show the both predictive power and feature selection ability of our methods.