Attribute Selection Effect on Tree-Based Classifiers for Letter Recognition

Abstract

This study presents evaluation measures for attribute selection effect on classification performance in classifying the 26 uppercase letters in the English alphabet. Attribute selection is an essential method in the classification phase to measure the attribute significance related to the class label since not all attributes are significant for letter recognition. Therefore, insignificant attributes should be reduced by applying dimensionality reduction. The filter-based attribute selection methods using Information Gain, Gain Ratio, Correlation, and Chi-square are proposed. The performances of attribute selection are evaluated by tree-based classifiers using J48, CART, and Random Forest algorithms with the measures of accuracy, precision, recall, F-measure, and processing time. The results indicate that the use of attribute selection methods provides the increase of classification performances for letter recognition. The reduction of insignificant attributes is discussed in terms of the effect on classification accuracy and the processing time. The optimal number of selected attributes is determined for each attribute selection, it provides better classification accuracy with more time-efficient.