A PCA Based Unsupervised Feature Selection Algorithm

Abstract

Principal components analysis (PCA) is an important approach to unsupervised dimensionality reduction. However, principal components (PCs) are a set of new variables carrying no clear physical meanings and still require all the original variables. To deal with this problem, the PC dominant feature (PCDF) is defined. Then, feature selection using them is considered and a new algorithm for determining such PC dominant features is proposed. Experimental results show that using the principal components as the basis the new algorithm can find a good feature subset.