Domains of competence of the semi-naive Bayesian network classifiers

The motivation for this paper comes from observing the recent tendency to assert that rather than a unique and globally superior classifier, there exist local winners. Hence, the proposal of new classifiers can be seen as an attempt to cover new areas of the complexity space of datasets, or even to compete with those previously assigned to others. Several complexity measures for supervised classification have been designed to define these areas. In this paper, we want to discover which type of datasets, defined by certain range values of the complexity measures for supervised classification, fits for some of the most well-known semi-naive Bayesian network classifiers. This study is carried out on continuous and discrete domains for naive Bayes and Averaged One-Dependence Estimators (AODE), which are two widely used incremental classifiers that provide some of the best trade-offs between error performance and efficiency. Furthermore, an automatic procedure to advise on the best semi-naive BNC to use for classification, based on the values of certain complexity measures, is proposed.