New models and algorithms for semi-naive Bayesian classification focused on the AODE paradigm.
Martínez Fernández, Ana María · 2012
Classification is one of the most popular tasks in machine learning, partly motivated by the high demand in real life applications, i.e.: spam filtering, cancer diagnosis, gene identification, credit assessment, etc. In this context, Bayesian networks entail an ideal framework to directly deal with the uncertainty encountered. In this dissertation, we work with the family of semi-naive Bayesian network classifiers (BNCs) that either do not perform structural learning or it is very simple. Particularly we focus on the Averaged One-Dependence Classifier (AODE), one of the most efficient and effective approaches to alleviate naive Bayes' independence assumption. This thesis comprises three main parts: ? The first one presents four new classifiers based on AODE. One of them is proposed to alleviate AODE's computational needs in terms of memory storage and classification time, which is called HODE (Hidden One-Dependence Estimators) . The other three are different alternatives to handle numeric attributes in BNCs: GAODE (Gaussian AODE) and HAODE (Hybrid AODE), based on (conditional) Gaussian networks; and MTE-AODE, based on the use of mixtures of truncated exponentials. ? The second one shows how the application of different disjoint (traditional) discretization techniques has a low impact on the performance of the semi-naive BNCs. Hence it studies and recommends the use of non-disjoint the discretization techniques as an attractive alternative. ? Experimental results indicate that the different alternatives proposed are beneficial for a particular group of datasets, whose common characteristics are unknown. That motivates the study of the domain of competence for these classifiers according to different complexity measures. Furthermore, an automatic procedure to advise on the best semi-naive BNC to use for a particular dataset is proposed.