Skip to main content
Back to timeline
Communications in computer and information scienceSource publication:

TabSCM turns tabular synthesis into per-variable causal equations, training up to 583 times faster than diffusion-only generators

Synopsis

TabSCM writes tabular data generation as a structural causal model: a causal discovery algorithm supplies a partially directed acyclic graph, the graph is oriented into a DAG, and each variable gets a conditional model that depends only on its causal parents, using conditional diffusion for continuous variables and gradient-boosted trees for categorical ones. On seven public datasets spanning healthcare, finance, housing and environment, it matches or surpasses GAN, diffusion and LLM baselines in statistical fidelity, downstream utility and privacy risk, while cutting rule-violation rates and running up to 583 times faster than diffusion-only models.

Source-provided article image: TabSCM: A Practical Framework for Generating Realistic Tabular Data
Fig. 1 · Page 2

Interpretation

Evaluation of tabular synthetic data has centered on marginal distributions, pairwise correlations and downstream task performance, none of which checks whether generated samples preserve causal relationships between variables, so downstream models can learn spurious associations. Earlier generators such as TabDDPM, TabSyn, TabDiff and GReaT target statistical fidelity or utility without explicitly constraining causal structure. Across seven public datasets, TabSCM matches or surpasses GAN, diffusion and LLM baselines in statistical fidelity, downstream utility and privacy risk, and has lower violation rates than diffusion-only models on three of four sanity checks.

Generation is decomposed into one equation per variable, each conditioned only on that variable's causal parents, so a single mechanism can be inspected or intervened on in isolation. Monolithic diffusion or transformer generators learn the joint distribution as one indivisible model and cannot localize an individual mechanism. On California Housing, the same gradient-boosted regressor trained on real data and on an equal-sized TabSCM synthetic set gives closely agreeing SHAP attributions on the same held-out real test instances, and mean absolute SHAP values for the MedHousVal and Latitude conditional models align with domain intuition.

Sampling variable by variable in topological order makes generated records satisfy known constraints by construction and supports exact counterfactual queries. Existing state-of-the-art models often produce out-of-domain entries that violate known constraints, limiting their use in high-stakes settings such as credit and healthcare. On Adult, Housing and Loan, four sanity checks show TabSCM with lower violation rates than diffusion-only models on three; on the California latitude-longitude boundary data, the real training data itself had a violation rate of 2.24%.

Perspective

The demonstrated scope is seven public tabular datasets covering classification and regression, with sample sizes from under one thousand to over 250,000, on tasks including income prediction, diabetes risk, credit default, loans, gamma-ray event classification, PM2.5 concentration and house prices; the method assumes causal sufficiency, meaning no unobserved confounders, and a data-generating process that is Markovian and faithful to a DAG.

Correlation-error quality is tied to the chosen causal discovery algorithm, and removing the modular decomposition can actually yield lower average correlation error; counterfactual explanations are only a preliminary analysis, with deeper analysis and extensions left to future work; diffusion steps and training epochs trade density error against correlation error.