Authors: Hari Seetha; M. Narasimha Murty; R. Saravanan
Addresses: School of Computing Science and Engineering, VIT University, Vellore-632014, India ' Department of Computer Science and Automation, Indian Institute of Science, Bangalore-12, India ' School of Information Technology and Engineering, VIT University, Vellore-632 014, India
Abstract: Text classification plays a vital role in the organisation of the unceasing growth of digital documents. High dimensionality of feature space is a major hassle in text classification. Feature selection, an effective preprocessing technique improves the computational efficiency and the accuracy of a text classifier. In the present paper, text classification is performed with Zipf's law-based feature selection and the use of linear SVM weight for feature ranking. A hybrid feature selection method combining these two feature selection techniques is proposed. Nearest neighbour and SVM classifiers are chosen as text classifiers for their good classification accuracy reported in many text classification tasks. Moreover, to investigate the effect of kernel type on the text classification both linear and non-linear kernels in SVM are examined. The performance is evaluated by determining classification accuracy using ten-fold cross-validation. Experimental results with four benchmark corpuses were encouraging and demonstrated that the classification performance using hybrid feature selection method outperformed the classification performance obtained by selecting either medium frequent features based on Zipf's law or using feature selection by linear SVM.
Keywords: text classification; SVM classifier; support vector machines; nearest neighbour classifier; feature selection; classification accuracy.
International Journal of Data Mining, Modelling and Management, 2015 Vol.7 No.3, pp.165 - 184
Available online: 28 Aug 2015 *Full-text access for editors Access for subscribers Purchase this article Comment on this article