Explorando estratégias bayesianas eficientes e eficazes para classificação de texto

Abstract

Automatic Document Classification (ADC) is the basis of many important applications such as spam filtering, opinion mining, content organizers and authorship identification. Naive Bayes (NB) approaches are widely used as a classification paradigm, due to their simplicity, efficiency and effectiveness in several scenarios. However, NB solutions do not present competitive effectiveness in Automatic Document Classification (ADC) tasks when compared to other modern statistical learning methods. In this dissertation, we investigate whether the combination of some alternative NB learning models with different feature weighting techniques can improve the NB effectiveness in ADC. We also present an investigation on the relaxation of the NB feature independence assumption (aka, Semi-Naive approaches) in large text collections. Given the high computational costs of these investigations, we present a massively GPU-based parallelized version of the NB. Moreover, supported by the parallel implementations, we propose four novel Lazy Semi-NB approaches. In our experiments, our solutions not only outperform existing Semi-NB approaches, but also surpass our improved NB solutions in terms of effectiveness that had already outperformed SVMs.