Background Dataset

The background dataset must accurately reflect the feature distribution of the production data. If the background sample is biased, the expected model output (baseline value) will be skewed, distorting all feature attributions.

• Purpose: Defines the 'average' prediction against which individual explanations are compared.
• Risk: Using a non-representative sample (e.g., only positive cases) invalidates the Efficiency Property in practice.
• Best Practice: Use a stratified random sample or cluster centroids (e.g., k-means) to capture the full data density.

The treatment of correlated features defines whether you compute Interventional SHAP or Observational SHAP. The background dataset's structure dictates this choice.

• Interventional: Breaks correlations by sampling from the marginal distribution. Requires a background dataset large enough to approximate independence.
• Observational: Preserves correlations by conditioning on known features. The background dataset must be dense enough to compute accurate Conditional Expectations.
• Impact: Interventional SHAP reflects a causal 'what-if' intervention, while Observational SHAP reflects the model's behavior on the natural data manifold.

SHAP explains a prediction by simulating the removal of features. The background dataset provides the values used to 'fill in' these removed features.

• Mechanism: When a feature is 'missing,' its value is replaced by a random draw from the background dataset's distribution for that feature.
• Consequence: If the background dataset contains outliers, these will be used as imputed values, leading to unrealistic model evaluations and noisy Shapley Value Estimation.
• Optimization: Clean the background dataset of extreme outliers to prevent the generation of out-of-distribution synthetic instances during explanation.

Exact Shapley value computation requires evaluating the model on all 2^M feature subsets, which is intractable. The background dataset's size directly impacts the convergence speed of sampling approximations.

• KernelSHAP: Uses a subset of the background dataset as a reference for its weighted linear regression. A smaller, well-chosen background set reduces computation.
• TreeSHAP: Uses the background dataset to compute internal node coverages for exact polynomial-time calculation.
• Variance Reduction: Techniques like paired sampling rely on a stable background distribution to reduce the number of model evaluations needed for convergence.

The Baseline Value is the expected model output over the entire background dataset. It serves as the starting point for every SHAP Waterfall Plot.

• Calculation: The average prediction when all features are 'missing' (i.e., replaced by background values).
• Interpretation: In a waterfall plot, the baseline is the initial value from which each feature's Marginal Contribution pushes the prediction up or down.
• Stability: A baseline computed from too few samples will have high variance, making explanations inconsistent across different runs.

For Causal SHAP, the background dataset must be compatible with the assumed causal graph. Standard SHAP assumes a flat, non-causal feature independence structure.

• Standard SHAP: Treats the background dataset as a bag of independent features, which can create unrealistic 'Frankenstein' instances when imputing.
• Causal SHAP: Requires the background dataset to be used in conjunction with a structural causal model to condition only on a feature's causal parents, not all features.
• Practical Rule: If features have strong physical or causal dependencies, a standard background dataset will produce explanations that are faithful to the model's function but not to the real-world data-generating process.