Stochastic gradient-CAViaR-based deep belief network for text categorization

Abstract

Text categorization is defined as the process of assigning tags to text according to its content. Some of the text classification approaches are document organization, spam email filtering, and news groupings. This paper introduces stochastic gradient-CAViaR-based deep belief networks for text categorization. The overall procedure of the proposed approach involves four steps, such as pre-processing, feature extraction, feature selection, and text categorization. At first, the pre-processing is carried out from the input data based on stemming, stop-word removal, and then, the feature extraction is performed using a vector space model. Once the extraction is done, the feature selection is carried out based on entropy. Subsequently, the selected features are given to the text categorization step. Here, the text categorization is done using the proposed SG-CAV-based deep belief networks (SG-CAV-based DBN). The proposed SG-CAV is used to train the DBN, which is designed by combining conditional autoregressive value at risk and stochastic gradient descent. The performance of the proposed SGCAV + DBN is evaluated based on the metrics, such as recall, precision, F-measure and accuracy. Also, the performance of the proposed method is compared with the existing methods, such as Naive Bayes, K-nearest neighbours, support vector machine, and deep belief network (DBN). From the analysis, it is depicted that the proposed SGCAV + DBN method achieves the maximal precision of 0.78, the maximal recall of 0.78, maximal F-measure of 0.78, and the maximal accuracy of 0.95. Among the existing methods, DBN achieves the maximum precision, recall, F-measure and accuracy, for 20 Newsgroup database and Reuter database. The performance of the proposed system is 10.98%, 11.54%, 11.538%, and 18.33% higher than the precision, recall, F-measure, and accuracy of the DBN for 20 Newsgroup database, and 2.38%, 2.38%, 2.37%, and 0.21% higher than the precision, recall, F-measure and accuracy of the DBN for Reuter database.