Bias reduction in skewed binary classification with Bayesian neural networks.

The Bayesian evidence framework has become a standard of good practice for neural network estimation of class conditional probabilities. In this approach the conditional probability is marginalised over the distribution of network weights, which is usually approximated by an analytical expression that moderates the network output towards the midrange. In this paper, it is shown that the network calibration is considerably improved by marginalising to the prior distribution. Moreover, marginalisation to the midrange can seriously bias the estimates of the conditional probabilities calculated from the evidence framework. This is especially the case in the modelling of censored data.