Self-supervised data synthesis using generative adversarial networks

The field of self-supervised learning has seen significant advancements, particularly in techniques that exploit unlabeled data. The present study does not introduce a new self-supervised pretext task; CTGAN is trained on an already labelled public table, and adversarial training is used as a generative (not labelling) objective. The current study investigates Conditional Tabular GAN (CTGAN), a tabular variant of Generative Adversarial Networks (GANs), for privacy-preserving synthetic data generation in a startup success prediction task. A two-module architecture: (i) a CTGAN-based generator that learns from the “Startup success prediction” Kaggle dataset and produces synthetic labelled tables, and (ii) an evaluation module, a stacked ensemble of classical classifiers, that quantifies how useful this synthetic data is for downstream prediction is proposed. Synthetic–real similarity is measured with column-level metrics: Column Shapes (85.73%), Column Pair Trends (74.54%), Data Quality (80.13%), KS Complement for continuous features (0.865), TV Complement for categorical features (0.908), Statistic Similarity (0.708), and Missing Value Similarity (1.0), indicating that CTGAN closely matches the marginal (column-wise) distributions of the original data. Column Pair Trends (74.54%) show that bivariate dependence is weaker, consistent with KS/TV measuring only univariate similarity. However, when nine classifiers are trained separately on real and synthetic data, the average cross-validation score is 0.73 on real data versus 0.54 on synthetic data, with the best real-data model (Random Forest) reaching 0.7818 compared to 0.5587 on synthetic data. These findings show that CTGAN is effective for high-fidelity, column-level tabular synthesis and useful as a complement to real data (e.g. mixed real–synthetic training) rather than as a stand-alone private replacement, but synthetic data alone is not yet a full replacement for real data in high-accuracy predictive modelling.