Improving Centroid-based Text Classification Using Term-distribution-based Weighting System and Clustering

Centroid-based text classification is one of the most popular supervised approaches to classify texts into a set of pre-defined classes with relatively low computation. Based on the vector-space model, the performance of this classification particularly depends on the way to weigh terms in documents in order to construct a representative class vector for each class and degree of spherical shape in class. In this paper, we propose a method to improve classification accuracy by considering a number of statistical term weighting systems based on term-distribution, including factors of intra-class, inter-class, overall term frequency distribution and term-length normalization. An improvement using a clustering technique called hierarchical EM is also investigated. A number of experiments using drug information web pages and newsgroups data set, are made. The results show that our method outperforms standard tf-idf centroid-based, k-nearest neighbor and naïve Bayesian classifiers to some extent.