Bayesian classification using a noninformative prior and mislabeled training data (Q5926827)

The performance of a Bayesian classification test is illustrated by an example in which the training data in each class are corrupted by mislabelling. Symbol probabilities are inferred by combining the information in the discrete training and the test data, these probabilities being assumed to have a Dirichlet prior distribution in each class. Using the average probability of error, the authors illustrate the effect that the mislabelling of training data has on the classification of unknown test data. The performance of a Bayesian data reduction algorithm in the improving of performance of mislabelled training data is examined.