Analysis of singular value decomposition using high dimensionality data

摘要

The aim of this paper is to analyze the sensitivity and accuracy of singular value decomposition (SVD) using high dimensionality data as database. Five standard dataset from the UCI Machine Learning Repository are utilized to evaluate and verify the significant principal components (PCs) using three rules of thumbs namely the Kaiser Gutman, Scree Test and Cumulative Variance. Upon identification of the PCs, these selected PCs are classified using Artificial Neural Network classifier. It was found that SVD as feature extraction improved the performance of classification accuracy and this is proven since recognition rate accuracy is higher with SVD as feature extraction as compared to original data solely. However, the inconsistency of identifying the number of significant PCs requires further research to be explored that might due to the effect of estimating the singular vectors as a whole instead of individually as well as the subspaces that these vectors span.