Abstract
Selecting features that represent a specific class is important to achieve a high text classification performance. The core and critical part of any text feature selection method is the weighting function. Most term weighting methods only consider document level when calculating a term weight and do not consider the distribution of features among different classes. Such an approach does not accurately reflect the specificity of each individual term that can discriminate between the positive and negative documents in the document collection because of the numerous uncertainties in text documents. To address this problem, we propose an innovative and effective feature-weighing method based on three-way decisions to reduce uncertainties in selected features. The proposed model can assign a more discriminately accurate weight to terms based on their distribution in each class. The experimental results, based on the standard RCV1 dataset and R21578 and three popular performance measures, show that our model significantly outperforms six state-of-the-art baseline models and enhances the performance of text classifier models.