The Effects of Globalization Functions on Feature Weighting for Text Classification

Text classification is crucial to easily access, index and storage by categorizing textual documents. An efficient categorization of text documents depends on the assignment of appropriate weights to the features besides using appropriate feature sets. This has attracted the researchers' attention to feature weighting methods for text classification. While some of the feature weighting methods in the literature generates a single global weight score for each feature, some of them generate class-based scores for each feature in the dataset. In this study, the impact of globalization functions on feature weighting for text classification is investigated in details. For this purpose, various experiments were carried out on 3 benchmark datasets using 3 different feature weighting methods, 2 different feature globalization functions, and 2 different classifiers. Also, various feature dimensions were used in the experiments in order to analyze the dependency between globalization functions and feature sizes. So, the impact of two different globalization functions have been tested for three different feature weighting methods namely TF.MI, TF.CHI2, and TF.PS. Experimental results obtained with SVM and KNN classifiers on the Reuters-21578, Mininew20, and WebKB datasets reveal that choosing appropriate globalization function for feature weighting methods may provide improvement on the performance of classification depending on various experimental settings used.