Semi-hierarchical naive Bayes classifier

The classification of high dimensional data is an arduous task especially with the emergence of high quality data acquisition techniques. This problem is accentuated when the whole set of features is needed to learn a classifier such as the case of genomic data. The Bayesian approach is suitable for these applications because it represents graphically and statistically the dependencies between the features. Unfortunately, learning a Bayesian classifier using a high number of features does not ensure a tradeoff between the dimensions' reduction, the semantic of the model and the predictive performance. We propose a new semi-hierarchical naive Bayes that uses the latent variables for abstracting the features of a given dataset in order to reduce the dimensionality. These variables are suitable for finding graphically and semantically analyzable models. We combined them with the observed variables in a tree-augmented naive Bayes structure in order to improve the prediction accuracy. An excessive experimental study showed that our method is suitable for high dimensional data and overcomes the existing methods.