EGAIN: Enhanced Generative Adversarial Networks for Imputing Missing Values

Missing data remain a persistent challenge in statistical analysis and machine learning because many predictive methods require complete observations. Generative Adversarial Imputation Networks (GAIN) offer a flexible deep-learning approach for missing value imputation, but their practical use is limited by convergence instability, sensitivity to hyperparameter selection, and dependence on outdated software implementations. To address these limitations, we propose Enhanced Generative Adversarial Imputation Networks (EGAIN), a modernized extension of GAIN implemented in TensorFlow 2.x. EGAIN incorporates convolution-based generator and discriminator networks, a channel-stacked representation of the data and mask, and checkpoint-based training diagnostics to improve stability and usability. EGAIN was evaluated on five benchmark datasets under multiple Missing Completely At Random (MCAR) settings and compared with the original GAIN implementation and median imputation. Across most evaluated conditions, EGAIN achieved lower root mean squared error (RMSE) and showed greater robustness, particularly when missingness was concentrated in a subset of variables. These results indicate that EGAIN provides a more stable and reproducible framework for missing data imputation in tabular datasets.