A-RDBOTE: an improved oversampling technique for imbalanced credit-scoring datasets

Banks and financial industries evaluate the creditworthiness of their customers through credit-scoring models before allocating loans to them. The performance of credit-scoring models significantly degrades due to class imbalance data, in which the class of defaulters is underrepresented as compared to that of non-defaulters, which is one of the major challenging tasks. In this paper, we propose a novel adaptive representative and density-based oversampling technique (A-RDBOTE) to deal with imbalanced credit-scoring datasets. First, the reverse k-nearest neighbor algorithm is applied to eliminate the noisy samples from the training set. Next, a semi-unsupervised clustering method is applied to cluster the minority instances. Then, from each sub-cluster, the representativeness of an instance is determined by considering its degree of similarity with respect to inter and intra-cluster. Subsequently, from each sub-cluster, the instances having high representative values are selected as anchor instances. Finally, artificial minority instances are generated around each anchor instance within the same sub-cluster. The experimental results showed that A-RDBOTE has achieved significantly better results than eight oversampling methods in terms of F1-score, AUC, and G-mean.