Empirical Evaluation of Centroid-based Models for Single-label Text Categorization

Abstract

Centroid-based models have been used in Text Categorization because, despite their computational simplicity, they show a robust behavior and good performance. In this paper we experimentally evaluate several centroidbased models on single-label text categorization tasks. We also analyze document length normalization and two different term weighting schemes. We show that: (1) Document length normalization is not always the best option in a classification task. (2) The traditional tfidf term weighting approach remains very effective, even when compared to more recent approaches. (3) Despite the fact that several ways to calculate the centroid of a class in a dataset have been proposed, there is one that always outperforms the others. (4) A computationally simple and fast centroid-based model can give results similar to the top-performing SVM model. 1 Introduction and Previous Work The main goal of text categorization (TC) is to derive models for the categorization of natural language text [19]. The objective is to derive models that, given a set of training documents with known categories and a new document, which is usually called the query, will predict the query’s category. Here, we are interested in the case