Feature Clustering based MIM for a New Feature Extraction Method

Abstract

In this paper, a new unsupervised Feature Extraction appoach is presented, which is based on feature clustering algorithm. Applying a divisive clustering algorithm, the method search for a compression of the information contained in the original set of features. It investigates the use of Mutual Information Maximization (MIM) to find appropriate transformation of clusterde features. Experiments on UCI datasets show that the proposed method often outperforms conventional unsupervised methods PCA and ICA from the point of view of classification accuracy.