Bayesian Sampling of Virtual Examples to Improve Classification Accuracy

A virtual example is an artificial example that does not exist in the given training set. We sample a virtual example from a Bayesian network constructed with the original training set. The usefulness of a sampled virtual example for learning is measured by the increment of the network´s conditional likelihood. A qualified virtual example is saved and used to update the network for the next sampling. By repeating this process we can generate candidate virtual example sets of various sizes. Among these candidates, an appropriately sized virtual example set for a target learning algorithm is chosen through statistical significance tests. Experiments have shown that the virtual examples collected this way can help various learning algorithms to derive classifiers of improved accuracy