Quantile kurtosis in ica and integrated feature extraction for classification

Abstract

As an effective statistic in independent component analysis (ICA), kurtosis can provide valuable information for testing normality, determining features shape and ordering independent components of feature extraction in classification analysis. However, it may lead to the poor performance in certain situations so that the quantile kurtosis has been developed. In this paper, we propose a robust quantile measure of kurtosis in ICA for feature extraction. Moreover, we also present a feature extraction method which integrates the extracted features of principal component analysis (PCA), linear discriminant analysis (LDA), ICA and random forest algorithm (RFA) together. For the ICA based feature extraction, independent components are sorted according to the proposed quantile kurtosis. The experimental results show that our integrated feature extraction method, especially with the help of the proposed quantile kurtosis, outperforms the others.