New Loss Functions Help Variational Autoencoders Generate Balanced and Fair Synthetic Data

An additive balance component integrated into a variational autoencoder training loss function successfully reduces sensitive feature imbalance and improves statistical parity in synthetic tabular data while maintaining utility and privacy at levels comparable to baseline models.

The research addresses operational risks surrounding training data quality, fairness, and privacy as machine learning and artificial intelligence systems are increasingly deployed in high-stakes environments. While synthetic data generation provides a recognized route toward mitigating these concerns, persistent class imbalance across sensitive categorical features continues to introduce unwanted bias into downstream models.

To steer generative modeling toward balanced distributions, the researchers formulated two alternative loss function mechanisms that penalize deviations from sensitive class balance: an additive balance component denoted as \(\mathcal{B}\mathcal{L}_{term}\) and a Simpson-index-based reweighting component designated as \(\mathcal{B}\mathcal{L}_{weight}\). Unlike existing methodologies in the published literature that require modifying the structural architecture of the variational autoencoder, these proposed formulations establish a lightweight and tunable means of controlling generation.

The experimental evaluation analyzed 1,260 synthetic datasets produced from three widely used benchmark datasets: Adult, South German Credit, and COMPAS. This empirical investigation systematically tested generation performance across two distinct batch-epoch configurations, seven loss function configurations, and thirty independent train and evaluation splits to measure trade-offs across balance, fairness, utility, and privacy.

The comparative findings demonstrated divergent behaviors between the two formulations across the evaluated benchmarks. While the additive balance component \(\mathcal{B}\mathcal{L}_{term}\) concurrently minimized sensitive feature imbalance and enhanced statistical parity without sacrificing data utility or privacy, the Simpson-index-based \(\mathcal{B}\mathcal{L}_{weight}\) reweighting approach showed a weaker impact on fairness metrics and degraded both utility and privacy when applied at high magnitudes.

Subject of Research: Synthetic tabular data generation and fairness in variational autoencoders

Article Title: Balanced synthetic data generation with variational AutoEncoders: a trade-off analysis of balance, fairness, utility and privacy

Article References: Rondina, M., Vetrò, A., Gilli, L., Kurapati, S., & De Martin, J. C. (2026). Balanced synthetic data generation with variational AutoEncoders: a trade-off analysis of balance, fairness, utility and privacy. Applied Intelligence, 56(16), Article 548. https://doi.org/10.1007/s10489-026-07532-3

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07532-3

Keywords: synthetic data, variational autoencoders, machine learning fairness, data privacy, tabular data, statistical parity, loss functions, data imbalance, Balanced, synthetic, data, generation