Abstract feature extraction for text classification

BİRİCİK G., DİRİ B., Sonmez A. C.

TURKISH JOURNAL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCES, vol.20, pp.1137-1159, 2012 (SCI-Expanded) identifier identifier


Feature selection and extraction are frequently used solutions to overcome the curse of dimensionality in text classification problems. We introduce an extraction method that summarizes the features of the document samples; where the new features aggregate information about how much evidence there is in a document, for each class. We project the high dimensional features of documents onto a new feature space having dimensions equal to the number of classes in order to form the abstract features. We test our method on 7 different text classification algorithms, with different classifier design approaches. We examine performances of the classifiers applied on standard text categorization test collections and show the enhancements achieved by applying our extraction method. We compare the classification performance results of our method with popular and well-known feature selection and feature extraction schemes. Results show that our summarizing abstract feature extraction method encouragingly enhances classification performances on most of the classifiers when compared with other methods.