Improved Relative Discriminative Criterion Feature Ranking Technique for Text Classification

Abstract

Feature ranking techniques are used to improve the performance of classification in text labeling problems. Most of the feature selection techniques utilize document and term frequencies to rank term. In contrast to document frequency, term frequency support real values of the term. Recent feature ranking techniques use term frequencies with frequently occurring terms, but ignore rarely occurring terms which are as meaningful and important as frequently occurring terms. Moreover, F-measure decreases as features of existing techniques increases. In this paper, Improved Relative Discriminative Criterion (IRDC) technique is proposed to obtain more informative and meaningful rarely occurring terms. IRDC scale up rarely occurring terms that is present in one class and absent in other classes. Additionally, IRDC creates a trade-off between frequently and rarely occurring terms. Experimental results indicate that our proposed technique on reuters21578 and 20newsgroup datasets using well known classifiers like multinomial naïve bayes (MNB), support vector machine (SVM) and decision tree (DT) performed better in terms of F-measure.