High Accuracy-priority Rule Extraction for Reconciling Accuracy and Interpretability in Credit Scoring

Abstract

Accuracy and interpretability are two perspectives that are difficult to balance; this is referred to as the accuracy-interpretability dilemma. If credit models gain interpretability, they lose accuracy, and vice versa. Researchers continue to develop an array of very complicated predictive models; however, the finance industry needs interpretable models that can be used in actual practice. Especially, advanced sequential ensembles are seldom considered in credit scoring. Therefore, it is worthwhile to explore new rule extraction methods capable of building sequential ensemble classifiers that are effective for credit scoring. To enhance the accuracy and interpretability of extracted rules, we extend continuous recursive-rule extraction (continuous Re-RX) to a high accuracy-priority rule extraction method referred to as continuous Re-RX with J48graft. Continuous Re-RX with J48graft uses a recursive approach called subdivision. This approach consists of a backpropagation neural network, pruning, and a J48graft decision tree for mixed datasets (those containing discrete and continuous attributes) to construct a high accuracy-priority rule extraction method. Compared with previous rule extraction methods for Australian- and German-based datasets, continuous Re-RX with J48graft achieved the highest accuracies, 88.4 and 79.0%, respectively, using tenfold cross validation (CV) and the Friedman and Bonferroni–Dunn tests, and 87.82 and 78.4%, respectively, using 10 runs of tenfold CV, with the best Friedman score. We also demonstrate how continuous Re-RX with J48graft overcomes the accuracy-interpretability dilemma based on its performance. We believe that continuous Re-RX with J48graft can help overcome the accuracy-interpretability dilemma for transparency of Big Data in financial situations and for industrial applications.