Resumen de Edited Naive Bayes
José María Martínez Otzeta, Basilio Sierra Araujo, Elena Lazkano Ortega, M. Ardaiz, Ekaitz Jauregi
Naive Bayes is a well-known and studied algorithm both in statistics and machine learning. Bayesian learning algorithms represent each concept with a single probabilistic summary. This paper presents a variant of the Naive Bayes method, in which the original training set is augmented in the following fashion: Leave-One-Out procedure is applied over the training set, and incorrectly classified instances according to Naive Bayes model are duplicated. The augmented dataset is used to induce the model. The motivation behind this idea is that giving more importance to hard instances (in this case, duplicating them) might contribute to make the model more accurate over that subset of the instance space. We have tested this algorithm over 41 UCI datasets. The results suggest that the chance of obtaining a significant better performance than with the original Naive Bayes approach are much greater than the opposite.
