Interpretable Robust Feature Selection via Joint -Norms Minimization

Dimension reduction is a hot topic in data processing field. The challenge lies in how to find a suitable feature subset in low-dimensional space to accurately summarize the important information in high-dimensional space, rather than redundant information or noise. This requires the proposed model to reasonably explain the importance of features and be robust to noise. In order to solve this problem, this paper proposes an interpretable robust feature selection method, in which both the reconstruction error term and the regularization term are constrained by -norm. The reconstruction error term can capture samples corroded by noise, while the regular term automatically finds a group of discriminative features on relatively clean samples. Experimental results show the effectiveness of the proposed method, especially on noise data sets.