JUCS - Journal of Universal Computer Science 31(7): 668-682, doi: 10.3897/jucs.124087
Meta-learning approach for variational autoencoder hyperparameter tuning
Michele Berti‡, Matheus Camilo da Silva‡, Sebastiano Saccani§, Sylvio Barbon Junior‡‡ University of Trieste, Trieste, Italy§ Aindo Srl, Trieste, Italy
Corresponding author: Matheus Camilo da Silva ( matheuscmilo@gmail.com ) © Michele Berti, Matheus Camilo da Silva, Sebastiano Saccani, Sylvio Barbon Junior. This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY-ND 4.0). This license allows reusers to copy and distribute the material in any medium or format in unadapted form only, and only so long as attribution is given to the creator. The license allows for commercial use. Citation:
Berti M, Camilo da Silva M, Saccani S, Barbon Junior S (2025) Meta-learning approach for variational autoencoder hyperparameter tuning. JUCS - Journal of Universal Computer Science 31(7): 668-682. https://doi.org/10.3897/jucs.124087 |  |
Abstract
Synthetic data generation is a promising alternative to traditional data anonymization, with Variational Autoencoders (VAEs) excelling at generating high-quality synthetic tabular datasets. However, VAE hyperparameter selection is often computationally expensive or suboptimal. We propose a meta-learning (MtL) method for hyperparameter recommendation, which achieves competitive performance to state-of-the-art Bayesian Optimization (BO) with median AUC values of 0.660 ± 0.038 (MtL) and 0.650 ± 0.041 (BO), showing no statistically significant difference. Notably, our approach reduces configuration time to under three minutes, compared to BO’s multi-hour requirement, while also enabling incremental improvements through new data integration. This combination of efficiency, adaptability, and performance establishes MtL as a practical solution for hyperparameter tuning in synthetic data generation.
Keywords
HPO, Meta-learning, Variational Autoencoders, VAE