Non-Disjoint Discretization for Aggregating One-Dependence Estimator Classifiers

There is still lack of clarity about the best manner in which to handle numeric attributes when applying Bayesian network classifiers. Discretization methods entail an unavoidable loss of information. Nonetheless, a number of studies have shown that appropriate discretization can outperform straightforward use of common, but often unrealistic parametric distribution (e.g. Gaussian). Previous studies have shown the Averaged One-Dependence Estimators (AODE) classifier and its variant Hybrid AODE (HAODE, which deals with numeric and discrete variables) to be robust towards the discretization method applied. However, all the discretization techniques taken into account so far formed nonoverlapping intervals for a numeric attribute. We argue that the idea of non-disjoint discretization, already justified in Naive Bayes classifiers, can also be profitably extended to AODE and HAODE, albeit with some variations; and our experimental results seem to support this hypothesis, specially for the latter.