Abstract
The use of synthetic data has become an increasingly relevant approach in machine learning, especially in situations involving limited, imbalanced, or sensitive datasets. This study presents a comparative analysis of two popular data generation methods—CTGAN and TVAE—in the context of both regression and classification tasks. Building on previous work focused on regression, this research expands to include classification by evaluating the impact of synthetic data on models such as Logistic Regression, Random Forest, and XGBoost, using the PIMA Indians Diabetes dataset as a case study. The primary focus is on how the number of training epochs affects the quality of the generated data and, consequently, the performance of downstream models. Results show that synthetic data can significantly improve classification accuracy, particularly in ensemble models, but the benefits are highly dependent on the generation method and training parameters. The study confirms the practical value of synthetic data while also highlighting the need for careful tuning to avoid overfitting and performance degradation.