Unsupervised feature selection based on principal components analysis

Abstract

An important issue related to mining large data sets, both in dimension and size, is of selecting a subset of the original features. In this thesis, we describe an unsupervised feature selection algorithm suitable for data sets, large in both dimension and size. The algorithm consists of two steps— Pre-selection and selection. Pre-selection is based on Procrustes Analysis, which keeps the original characters as many as possible. The second step is based on feature similarity measure, with the aim of reducing the feature redundancy.