TabSCM: A practical Framework for Generating Realistic Tabular Data

Technical Summary: TabSCM – A Practical Framework for Generating Realistic Tabular Data

1. Problem Statement

Generating high-quality synthetic tabular data is a significant challenge due to the presence of heterogeneous data types (continuous and categorical), non-linear dependencies, and complex causal relationships. Existing state-of-the-art (SOTA) generative models—such as GANs, Diffusion models (e.g., TabDDPM), and Large Language Models (e.g., GReaT)—primarily focus on matching marginal statistics (individual column distributions) or pairwise correlations.

However, these methods are often causally agnostic. By ignoring the underlying causal structure, they may:

• Learn spurious correlations that do not exist in reality.
• Generate "semantically invalid" samples (e.g., an age that contradicts years of work experience).
• Fail to support valid counterfactual reasoning or policy simulations.
• Be computationally expensive (in the case of pure diffusion models).

2. Methodology

The authors propose TabSCM, a framework that leverages Structural Causal Models (SCMs) to ensure that synthetic data respects the causal dependencies of the original dataset. The process follows four main stages:

1. Causal Discovery & Graph Refinement: The framework starts with a Completed Partially Directed Acyclic Graph (CPDAG) obtained via causal discovery algorithms (e.g., PC, GES, or NOTEARS). It then orients the edges to form a valid Directed Acyclic Graph (DAG).
2. Topological Decomposition: Instead of modeling the joint distribution of all variables at once, TabSCM decomposes the joint distribution into a set of structural assignments (conditional distributions) based on a topological ordering of the nodes.
3. Hybrid Conditional Modeling: The framework uses specialized models for different variable types:
   • Root Nodes: Marginal distributions are estimated using Kernel Density Estimation (KDE) for continuous variables and relative frequencies for categorical ones.
   • Continuous Child Nodes: Modeled using Conditional Denoising Diffusion Probabilistic Models (DDPM), which learn to reconstruct the variable conditioned on its causal parents.
   • Categorical Child Nodes: Modeled using Gradient-Boosted Trees (decision trees) to provide a robust and flexible approximation of conditional probabilities.
4. Ancestral Sampling & Intervention: Data is generated by following the topological order. Because the model is an SCM, it allows for exact counterfactual queries using the do-operator. One can intervene on a specific variable (fixing its value) and propagate that change through its descendants to simulate "what-if" scenarios.

3. Key Contributions

• Causal-Generative Integration: A novel way to combine the high fidelity of diffusion models with the structural rigor of SCMs.
• Computational Efficiency: By decomposing the problem into per-variable equations, the model is up to 583× faster to train than pure diffusion-only models.
• Semantic Validity: The topological generation process inherently reduces rule violations (e.g., ensuring age and experience remain logically consistent).
• Mechanism-Level Interpretability: Unlike "black-box" generators, TabSCM allows users to inspect, debug, and audit individual causal mechanisms (the conditional models for each variable).
• Imbalanced Learning Support: Demonstrates that conditional generation is a principled alternative to heuristic upsampling (like SMOTE) for improving minority class fairness.

4. Experimental Results

The framework was evaluated on seven diverse datasets (healthcare, finance, housing, etc.) against baselines like CTGAN, TabDDPM, TabSyn, and GReaT.

• Statistical Fidelity: TabSCM matches or surpasses SOTA models in density estimation and correlation error.
• Downstream Utility: It achieves competitive AUC/RMSE scores, proving that models trained on TabSCM data perform well on real-world tasks.
• Privacy: It maintains a high Distance to Closest Records (DCR), ensuring that synthetic samples do not simply "copy" real individuals.
• Validity: It significantly outperforms other models in "sanity checks" (e.g., ensuring geographic coordinates match specific states or that marital status aligns with gender).
• Robustness in Imbalance: In imbalanced settings (e.g., HELOC credit risk), TabSCM improved minority fairness (lower False Negative Ratio) and trustworthiness (lower False Positive Ratio).

5. Significance

TabSCM represents a shift from purely statistical data synthesis to causally-aware synthesis. Its significance lies in its practicality for high-stakes, regulated domains (finance, medicine, law). By providing a tool that is simultaneously fast, interpretable, and causally sound, it enables researchers and policymakers to perform reliable simulations, audit models for fairness, and generate high-utility datasets that respect the fundamental laws of the data-generating process.