AI / ML
Synthetic Data Generation (M.S. Thesis)
Generative models combining VAEs, GANs and kernel density estimation to produce high-fidelity synthetic tabular data.
- 99%accuracy
- 15dimensions
- 19%faster training (AIENs)
Problem
Real data is often scarce, sensitive or locked behind privacy rules. Synthetic data can stand in for it, but only if it keeps the statistical structure of the original.
Approach
- Combined GANs, CTGANs, VAEs, KDE and multivariate Gaussian distributions, and verified the outputs with K-means and Gaussian Mixture Model clustering.
- AIENs (Adaptive Initializer Enhancer Networks): used adaptive kernel density estimation in place of the Gaussian prior in the VAE’s KL-divergence loss. Training time dropped by 19% because the prior already captures the data pattern, and data fidelity improved by 9%.
- Now extending the work to matrix manifolds, so generation respects the data’s underlying geometry.
Results
Generated synthetic data up to 15 dimensions at up to 99% accuracy. Deployed as a Django web app on AWS: generated data is saved to S3 through Lambda, and code ships from GitHub to EC2 with CodeDeploy.
Advisors: Prof. Michael Soltys and Dr. Ricardo Suarez.