Improving Classification Accuracy of Automated Text Classifiers

Abstract

The number of textual documents is increasing at an incredible rate and very often, there is a need to classify these documents into some fixed predefined categories automatically. Since the classification is being done automatically, the classifier needs to be a good classifier so that there are as less misclassifications as possible. Therefore, the classification accuracy is very important and needs to be taken care of. There are various factors that can affect the classification accuracy of classifiers. One of the factors is the Feature Selection method used to reduce the number of features in the documents. Information Gain (IG) is one of the most popular measures employed for this task but there are few shortcomings in this measure of evaluating the useful words for classification. In this paper, a new measure is proposed which eliminates those shortcomings and thus finds the better words which are more useful in the classification task. The new proposed measure aims to find those words which have more discriminating power than others and therefore, it is named as Discriminating Power (DP). We have also compared the results of using IG measure and DP measure for text classification and the results show that new proposed measure takes almost same time as IG measure but still improves the average classification accuracy of a text classifier and is much more consistent or stable in its classification accuracy when count of features to be selected is varied.