Neural network for ordinal classification of imbalanced data by minimizing a Bayesian cost

Ordinal classification of imbalanced data is a challenging problem that appears in many real world applications. The challenge to simultaneously consider the order classes and class imbalance, which can notably improve performance metrics. Bayesian formulation allows deal with these two characteristics jointly: It takes into account prior probability each decision costs, be used include imbalance ordinal information, respectively. We propose use train neural networks, have shown excellent results tasks. A loss function proposed networks single neuron output layer threshold based rule. an estimate cost, on Parzen windows estimator, fitted for thresholded decision. Experiments several datasets show method provides competitive different scenarios, due its high flexibility specify relative importance errors patterns classes, considering independently class.