Interactions between document representation and feature selection in text categorization

Many studies in automated Text Categorization focus on the performance of classifiers, with or without considering feature selection methods, but almost as a rule taking into account just one document representation. Only relatively recently did detailed studies on the impact of various document representations step into the spotlight, showing that there may be statistically significant differences in classifier performance even among variations of the classical bag-of-words model. This paper examines the relationship between the idf transform and several widely used feature selection methods, in the context of Naive Bayes and Support Vector Machines classifiers, on datasets extracted from the dmoz ontology of Web-page descriptions. The described experimental study shows that the idf transform considerably effects the distribution of classification performance over feature selection reduction rates, and offers an evaluation method which permits the discovery of relationships between different document representations and feature selection methods which is independent of absolute differences in classification performance.