Effective Learning with Joint Discriminative and Representative Feature Selection

Feature selection plays an important role in various machine learning tasks such as classification. In this paper, we focus on both discriminative and representative abilities of the features, and propose a novel feature selection method with joint exploration on both labeled and unlabeled data. In particular, we implement discriminative feature selection to extract the features that can best reveal the underlying classification labels, and develop representative feature selection to obtain the features with optimal self-expressive performance. Both methods are formulated as joint \( \ell_{2,1} \)-norm minimization problems. An effective alternate minimization algorithm is also introduced with analytic solutions in a column-by-column manner. Extensive experiments on various classification tasks demonstrate the advantage of the proposed method over several state-of-the-art methods.