A Semi-supervised Method for Feature Selection

Feature selection has been widely studied in the literature in both supervised and unsupervised scenario for dimensionality reduction. Supervised methods may cost too much on labeling, while unsupervised ones may lose efficacy because of lack of labels. In order to reduce dimensionality with less expense and higher efficiency, we propose a novel semi-supervised method based on Linear Discriminant Feature Selection (LDFS) and graph optimization framework, called Semi-supervised Discriminant Feature Selection (SDFS), which makes use of both labeled and unlabeled samples. Specifically, a small number of labeled data points are used to maximize the separability between different classes and a large amount of unlabeled data points are used to estimate the intrinsic geometric structure of the data. Experiments of dimensionality reduction show that our new feature selection methods out-perform related state-of-the-art feature selection approaches. SDFS utilizes both discriminant structure information of the labeled samples and geometric structure information of the unlabeled samples, which can improve the efficiency of dimensionality reduction in condition of only a few labels.