A replication of a cost-sensitive decision tree approach for fraud detection

Abstract

Payment card fraud poses a problem with widespread implications. However, the class imbalance between genuine and fraudulent transactions presents a challenge to traditional learning methods. Cost-based methods address this problem by assigning different costs to the misclassification of each class. Sahin et al. [7] define a cost-sensitive decision tree algorithm, and this study attempts to both replicate and build on their findings. The results presented here do not support the hypothesis that the cost-sensitive algorithm provides an improvement over traditional decision tree algorithms. However, the results do suggest that different methods of prioritizing alerted transactions can lead to performance improvements over some measures.