Kernel Paired Copula Bayesian Classifier

Resumo:

Bayesian networks, also known as causal networks, belief systems, or probabilistic dependency plots emerged in the 1980s and were applied in a wide variety of real-world activities. However, the most common available structure estimation algorithms underlying Bayesian network classifiers are often confined to discrete or Gaussian models. Recently, Copula methodology is considered to handle classification with Bayesian networks such as Copula Network Classifiers (CNC). This article compares the CNC with a Bayesian network of copula pairs (PCCBN) considering bivariate classification problems. The predictive performance of the both methods are compared in the modeling of real and artificial datasets. Furthermore, we propose a new Bayesian classifier based on non-parametric paired copula construction (KPCCBN).