ECS-SMOTE: A New Over-Sampling Method for Example-Dependent Costs Classification

Many real-world decision problems can fundamentally be categorized as classification tasks, where predictive models play an increasingly important role in optimizing decision-making and minimizing costs, such as in finance and health. Thus, Example-dependent Cost-sensitive Classification garners significant attention. We have extended the existing SMOTE method by incorporating example-dependent misclassification costs into the process of generating synthetic samples, leading to the proposal of the ECS-SMOTE method. We evaluate the proposed method on six real-world datasets (in the domains of credit scoring, direct marketing, and churn modeling). The experiments show that our method outperforms three classical oversampling methods and five EDCS classification models in terms of cost-sensitive metrics (such as Savings).