Abstract: | This thesis compares rough sets theory (RS) and principal component analysis (PCA) used for preprocessing coarsely discrete numerical datasets for classifiers based on neural networks (NN). Preprocessing the training data will often improve the learning time of NN. For numerical datasets, PCA has a long proven record of being a competent front-end to NN. Recently, RS has been suggested for reducing the dimensionality of nominal datasets. Many datasets are characterised by having only coarsely discrete numerical values, and both PCA and RS are potential candidates for reducing such datasets. The result of the experiments in this thesis shows that, for the particular classification problems studied, RS is a competent alternative to PCA as a front-end to NN when the dataset contains coarsely discrete numerical values.
|