A Study of Data Reduction Using Multiset Decision Tables

Abstract

Automatic text classification has gained huge popularity with the advancement of information technology. Bayesian method has been found highly appropriate for text classification but it suffers from a number of problems. When there is large number of categories, lack of uniformity in training data becomes a big problem. Some nodes may get less training documents, while other may get a very large number. Therefore, some nodes are biased over others. Besides, presence of noise data or outliers also creates problems. Moreover, when documents are very small, just like a line item describing a product, the problem becomes more difficult. In this paper we describe a method that combines naive Bayesian text classification technique and neural networks to handle these problems. We start with a naive Bayesian classifier, which has the linear separating surfaces. We modify the separating surfaces using neural network to find better separating surfaces and hence better classification accuracy over validation data.