HySMOTE: A Hybrid Approach for Handling Class Imbalances*

Synthetic data generation provides a feasible solution to handle class imbalances in real-world datasets. To generate diverse and information-rich synthetic samples, it is important to consider both the majority and minority class samples, as well as data complexities in the generation process. Therefore, we proposed a hybrid approach that combines traditional SMOTE with a counterfactual generation method while exploring three data complexities. To achieve this, the dataset is first divided into safe, disjuncts, borderline and outlier regions using clustering. The outliers are removed to avoid their use in the data generation process. Next, to balance the dataset, minority class samples are generated in the safe region using the traditional SMOTE technique, while samples in the borderline and disjunct regions are generated through the counterfactual generation method. The results show that the ranking of our …