Publication: An analysis of two dimensionality reduction techniques on the performance of neural network classifiers
Date
2005-03-01
Authors
Ong, Siok Lan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This project involves an analysis of the effectiveness of two dimensionality reduction techniques, i.e., Principal Component Analysis as the standard approach and Random Projection as a recent technique. The study is based on the performance of two supervised neural network classifiers i.e., Standard Backpropagation and Fuzzy ARTMAP. A set of benchmark and real medical databases are used to evaluate the performance of the neural network models. The performance indicators used are percentage of correct classification, purity, and collective entropy. The Student’s two-tailed paired t-test is used to compare the significance of differences of the results. Based on the estimated 95% confidence intervals, a strong decision which eventually leads to a convincing conclusion on the performance of the dimensionality reduction techniques can be obtained. The perceived experimental results especially from the real medical data sets are encouraging enough to prove that Random Projection exhibits good performance as a dimensionality reduction technique. Surprisingly, Random Projection is effective on low dimensional data, and the outcomes are as good as Principal Component Analysis. A discussion on generalization of the results obtained is included, and a conclusion ensues. Recommendations are also included for further improvements and enhancements in the analysis of dimensionality reduction techniques.