Categorical term descriptor: a proposed term weighting scheme for feature selection

Abstract

This paper proposes a term weighting scheme, categorical term descriptor (CTD), for feature selection in automated text categorization. CTD is an adaptation of the term frequency inverse document frequency (TFIDF). We compared the performance of the proposed method against classical methods such as correlation coefficient, chi-square and information gain using the multinomial naive Bayes and the support vector machine (SVKD) classifiers on the Reuters(10) and Reuters (115) variants of Reuters-21578 dataset. Despite its simplicity, CTD has proven to be promising for both local and global feature selection. CTD works best for the Reuter(10) as a stable local FS method.