A K-means Improved CTGAN Oversampling Method for Data Imbalance Problem

CTGAN is a tabular data synthesis method for privacy preservation, which is used in this paper for data imbalance problem. This paper proposes a method for dealing with imbalanced data sets that combines K-means clustering and CTGAN to address the imbalanced distribution of minority class examples that result from oversampling with CTGAN. By conducting experiments with the LightGBM algorithm on home loan and online shopping datasets, it is demonstrated that the CTGAN method achieves superior learning results in f1-score and G-mean metrics compared to the interpolation-based oversampling technique represented by SMOTE. The preceding results indicate that by applying the method described in this paper to handle an imbalanced dataset, one can obtain a dataset with more examples, a more uniform distribution, and less overfitting while still satisfying the original dataset's probability distribution.