Toward an Auto-Synthesizer: Stress-Aware Selection and Recommendation of Deep Tabular Generative Models

Deep generative models for tabular data — GANs, variational autoencoders, diffusion models, and LLM-based generators — exhibit highly non-uniform behavior across datasets: the best-performing family depends on distributional stressors such as long-tailed marginals, high-cardinality categoricals, Zipfian imbalance, and small-sample regimes. While automated machine learning resolved the analogous selection problem for supervised learning a decade ago, tabular synthesis still lacks an “Auto-Synthesizer”: its libraries standardize and benchmark generators but leave the question of which model to use to the practitioner, for whom evaluating every candidate — training, sampling, and scoring each under competing objectives of fidelity, privacy, and utility — is prohibitively expensive. This thesis studies intent-conditioned tabular synthesis selection and answers it with a two-system framework. Table-Synthesizers, the generation and measurement substrate, unifies 19 published generators behind a single interface and manufactures the empirical ground truth against which any selector must be judged. SYNTHONY, the selection layer, introduces stress profiling — a synthesis-specific, interpretable meta-feature representation quantifying dataset difficulty along four failure-mode-aligned dimensions — and matches stress profiles against a measurement-grounded, Bayesian-calibrated capability registry to rank candidates without training a single generator at inference time. Across a benchmark of 7 datasets, 10 synthesizers, and 3 user intents, stress-based meta-features prove highly predictive of synthesizer performance: an oracle-informed nearest-neighbor reference attains 0.578 held-out Spearman rank correlation against intent-specific oracles, which this thesis independently reproduces (0.581) from the recovered evaluation matrix. The deployable zero-shot selector attains a reported 0.573 — within 0.005 of that reference at nine held-out pairs, a gap far smaller than the coarse granularity the sample size supports — while training no generator at inference time, remaining per-dimension inspectable and extensible to newly registered generators without benchmark reruns; independently reproducing that figure awaits recovery of the selector’s calibrated parameters from the original pipeline. Ablations show dataset stress profiling is the load-bearing component, and that freeing all registry parameters overfits at current ground-truth scale — motivating the framework’s re-runnable calibration pipeline. The reproduced selection signal is robust to the definition of the third intent, holding (Spearman 0.52–0.58) whether downstream quality is measured by latency or a train-on-synthetic–test-on-real utility oracle. The thesis closes by specifying, at task granularity, the remaining distance to a closed-loop Auto-Synthesizer that executes, evaluates, and learns from its own recommendations.