SAMPLE SELECTION AND NEURAL NETWORK RULE EXTRACTION FOR CREDIT SCORING

摘要

We present an approach for sample selection using an ensemble of neural networks(NNs) for credit scoring. The ensemble determines samples that are outliers by checkingthe NN prediction accuracy on the original training data samples. Samples that areconsistently misclassied by NNs in the ensemble are removed from the training dataset. The remaining data samples are used to train another NN for rule extraction.Our experimental results show that by eliminating the outliers, NNs can be trained toachieve better predictive accuracy. The rule set extracted from one of these networksis more accurate than the rule set extracted from NNs trained with the original data.