A Regularization Framework for Feature Selection

摘要

This paper presents a framework aiming to find the relevant features using both the labeled and unlabeled data. Within a weighted space the discriminant structure of the data set is inferred by the labeled data points, and the intrinsic geometrical structure of the data set is inferred by the mixed labeled and unlabeled data points. From the framework we derive two feature selection algorithms, i.e. Semi-supervised Feature Ranking by Linear Discriminant Analysis(SFRLDA) and Semi-supervised Feature Ranking by Discriminant Neighborhood Analysis(SFRDNE). A series of experiments show that the proposed approaches can outperform previous methods in terms of the test accuracy on the synthetic and real-world benchmark data sets.