Introduction
The rapid adoption of "black-box" machine learning models has fueled the field of eXplainable AI (XAI). However, a fundamental crisis exists: we lack ground truth for explanations on tabular data. Unlike computer vision, where we can intuitively judge if a model is looking at the right pixels, tabular data relationships are abstract and often highly correlated. When tools like SHAP or LIME provide attribution scores, there is no objective way to know if they reflect the model's true logic or are misled by data artifacts. This paper introduces a framework for generating synthetic tabular datasets where the "ground truth" for feature importance is mathematically defined.
Methodology: A Three-Part Framework
The authors propose a synthetic data generation process that combines statistical control with symbolic rigor:
- Copula-Based Feature Generation: By using copula theory, the framework separates marginal distributions from correlation structures. This allows for the creation of features with specific statistical properties (e.g., Gaussian or Uniform marginals) while precisely controlling inter-feature correlations.
- Feature Categorization: Features are explicitly defined as Informative (used in the symbolic expression), Redundant (linear combinations of informative features), or Nuisance (pure noise).
- Symbolic Labeling: The relationship between informative features and the output is defined using symbolic math expressions. This regression value is passed through a sigmoid function to generate binary classification labels.
Key Insights into XAI Behavior
By controlling the underlying data-generating process, the authors identified several critical ways in which XAI tools can be misled:
- Correlation Artifacts: Correlation between informative features can "smear" local attributions. An explainer might report lower importance for a feature simply because it is correlated with another, an artifact of how explainers calculate expectations over the training data distribution rather than a reflection of the model's learned logic.
- The Nuisance Baseline: Including nuisance features (pure noise) establishes a floor for explainability. If an XAI tool assigns significant importance to these features, it indicates a failure in the explanation method.
- Redundancy Distortion: Adding redundant features significantly impacts SHAP's perception of importance. Summing the attributions of redundant features often fails to recover the original baseline importance, making local attributions difficult to interpret.
Results and Evaluation
The framework was used to evaluate SHAP (KernelExplainer and DeepExplainer) against synthetic ground truth. While SHAP correctly identifies dominant features in simple 1D cases, its performance degrades as complexity increases.
In experiments with correlated and redundant variables, the authors proved that SHAP values often reflect the data distribution's artifacts rather than the symbolic ground truth. This highlights the necessity of using controlled synthetic benchmarks to validate XAI methods before deploying them in high-stakes tabular domains like credit scoring or medical diagnostics.
Conclusion
This work provides a rigorous path forward for the XAI community. By moving away from subjective evaluation and toward ground-truth-based benchmarking, researchers can better understand the limitations of current tools and develop more robust explanation methods for tabular data.
Summary updated