Inducing Valuable Rules from Imbalanced Data: The Case of an Iranian Bank Export Loans

Credit scoring is a classification problem leading to introducing numeroustechniques to deal with itsuch as support vector machines, neural networks and rule-based classifiers. Rule bases are the top priority in credit decision making because of their ability to explicitly distinguish between good and bad applicants. In a creditscoring context, imbalanced data sets frequently occur as the number of good loans in a portfolio, which is usually much higher than the number of loans that default.The paper is to explore the suitability of RIPPER, OneR, Decision table, PART and C4.5 for loandefault prediction rule extraction. A real database of one of Iranian banks export loans is used, and class imbalance issues are investigated in its loan database by random oversampling the minority class of defaulters along with three sampling of majority in nondefaulters class. The performance criterion chosen to measure such an effect isthe area under the receiver operating characteristic curve (AUC), accuracy measure and number of rules. Friedman’s statistic is used to test significant differences between techniques and datasets. Theresults shows that PART is the best classifier in all of balanced and imbalanced datasets.