Discriminative Semi-Supervised Feature Selection Via a Class-Credible Pseudo-Label Learning Framework

Abstract

Most of existing semi-supervised learning methods heavily depend on labeled samples and always indistinguishably regard all unlabeled instances. However, unreliable samples lying around the boundary lines severely disturb training models. To address the problems, we pose a Class-credible Pseudo-label Learning (CPL) framework for semi-supervised data analysis. CPL is a classification model by optimizing the pseudo-label matrix which is probabilistic and exponentially controlled by the coefficient γ. By virtue of it, the model can identify the poor samples who have indistinguishable classes and enhance the solid samples with high class-credibility. That means class-indeterminate samples could be weakened, while class-definite ones could be enhanced. We then present a concise discriminative feature selection model with ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2,p</inf>-norm (p ∈ (0, 1)) regularization. Extensive experiments on several datasets demonstrate the superior performance of proposed method against representative competitors.