A Feature Selection Method for Improved Document Classification

Abstract

The aim of text document classification is to automatically group a document to a predefined class. The main problem of document classification is high dimensionality and sparsity of the data matrix. A new feature selection technique using the google distance have been proposed in this article to effectively obtain a feature subset which improves the classification accuracy. Normalized google distance can automatically extract the meaning of terms from the world wide web. It utilizes the advantage of number of hits returned by the google search engine to compute the semantic relation between two terms. In the proposed approach, only the distance function of google distance is used to develop a relation between a feature and a class for document classification and it is independent of google search results. Every feature will generate a score based on their relation with all the classes and then all the features will be ranked accordingly. The experimental results are presented using knn classifier on several TREC and Reuter data sets. Precision, recall, f-measure and classification accuracy are used to analyze the results. The proposed method is compared with four other feature selection methods for document classification, document frequency thresholding, information gain, mutual information and χ ² statistic. The empirical studies have shown that the proposed method have effectively done feature selection in most of the cases with either an improvement or no change of classification accuracy.