Feature Space Reductions Using Rough Sets for a Rough-Neuro Hybrid Approach Based Pattern Classifier
Ashwin Kothari, Avinash G. Keskar, S Srinath, Rakesh Chalasani · 2008
Abstract—Use of rough set theory at preprocessing stage for the targeted dimensionality reduction is very important in case of Unsupervised ANN based pattern classifiers. Such attribute compression by removing the redundancies results in reduced set of attributes acting as the input for the unsupervised neural network, based on the Kohonen’s learning rule. This would result in speeding up the learning process, better understanding about the data by recognizing the significant contributors to discernibility. Here the case of classification of printed alphabetical characters is taken, for which letters A-Z of eighteen different fonts form the data set. The features extracted are broadly classified into statistical and structural categories. The statistical features are inconsistent compared to the structural features. Exploiting inconsistencies of the data by rough sets has been observed and which results in exclusion of some important structural features. Inclusion of such features results in improved classification by the ANN.