Optimisation of Bayesian Classifiers by Using a Splitting Hierarchical EM Algorithm

The most important feature of a classifier is its generalisation capability. It depends on the correct estimate of both the parameters and the structural complexity of the network. When dealing with Bayesian classifiers, such a problem is usually related to the optimisation of different instances of the EM algorithm, which is used in the Maximum Likelihood approach. At this regard, we propose an EM-based algorithm that performs this optimisation task with a relevant reduction of the computational cost without loss of the classification accuracy. This is obtained by using a hierarchical growing approach, based on a given splitting procedure, which determines in a more efficient way the overall structural complexity of the classifier and thus its generalisation capability.