An empirical comparison of dimensionality reduction techniques for pattern classification

Abstract

To some extent or other all classifiers are subject to the curse of dimensionality. Consequently, pattern classification is often preceded with finding a reduced dimensional representation of the patterns. In this paper we empirically compare the performance of unsupervised and supervised dimensionality reduction techniques. The data set we consider is obtained by segmenting cells in cytological preparations and extracting 9 features from each of the cells. We evaluate the performance of 4 dimensionality reduction techniques (2 unsupervised) and (2 supervised) with and without noise. The unsupervised techniques include principal component analysis and self-organizing feature maps, while the supervised techniques include Fisher's linear discriminants and multi-layered feed-forward neural networks. Our results on a real world data set indicate that multi-layered feed-forward neural networks outperform the other three dimensionality reduction techniques and that all techniques are sensitive to noise.