Tractable Bayesian Learning of Tree Augmented Naive Bayes Models

Bayesian classiers such as Naive Bayes or Tree Augmented Naive Bayes (TAN) have shown excellent performance given their sim-plicity and heavy underlying independence assumptions. In this paper we introduce a classi er taking as basis the TAN model and taking into account uncertainty in model se-lection. To do this we introduce decompos-able distributions over TANs and show that they allow the expression resulting from the Bayesian model averaging of TAN models to be integrated into closed form. With this re-sult we construct a classier with a shorter learning time and a longer classication time than TAN. Empirical results show that the classi er is, most of the cases, more accurate than TAN and approximates better the class probabilities. 1