A New Gaussian Clustering Method for High Dimensional Classification Problems

Abstract

We propose a new Gaussian clustering method named EM-FDA for feature extraction in high dimensional classification problems. In this method, the distribution of each class is approximated as a mixture of Gaussians, which is clustered by applying Expectation-Maximization (EM) algorithm in the lower-dimensional Fisher’s discriminant space. Compared with conventional EM algorithm and other Gaussian clustering models, we show that the new method is more adaptable to various data distribution types and robust for widely ranging training sample sizes. Extensive experiments and comparison results with synthetic data, benchmark datasets and a real computational vision problem are presented.