Abstract

The growing adoption of ML/AI systems in high-stakes contexts has raised concerns about the quality, fairness and privacy of training data. Although synthetic data offers a promising way to address these concerns, the imbalance of sensitive categorical features still represents a source of bias. This paper presents two alternative approaches for a Variational AutoEncoder (VAE) training loss function that penalise deviations from a balanced distribution of sensitive feature classes. These are an additive balance component (BLterm) and a Simpson-index-based reweighting component (BLweight). In comparison with other approaches in literature, which alter the architecture of the VAE, our method is a lightweight and tunable mechanism to steer the generation towards balance. An empirical evaluation was conducted on 1,260 synthetic datasets generated from three widely used benchmark datasets (Adult, South German Credit, and COMPAS), across two batch-epoch configurations, seven loss function configurations and thirty train/evaluation splits. The results show that BLterm reduces sensitive feature imbalance and improves statistical parity, while preserving utility and privacy at levels comparable to the baseline. The BLweight formulation, although capable of reducing imbalance, exhibits a weaker impact on fairness and may degrade utility and privacy at high magnitudes. The proposed contribution offers practitioners a simple, tunable mechanism for improving the balance and fairness of synthetic tabular data generation without compromising utility or privacy.