Global Surrogate Model

The core principle is substituting an opaque model with a transparent one for explanation purposes. The surrogate is trained on a dataset where the input features are the original data and the target labels are the predictions of the black-box model.

• Common surrogates: Shallow decision trees, linear regression, logistic regression, or Explainable Boosting Machines (EBMs).
• Fidelity metric: The surrogate's accuracy is measured by how well its predictions match the black-box model's predictions (R-squared or classification accuracy), not the ground truth.

The surrogate model is trained using a two-step process that decouples explanation from original training.

• Step 1: Generate a transfer dataset. Use the original input data, pass it through the black-box model, and collect the output predictions.
• Step 2: Train the white-box model. Fit the interpretable surrogate model on this new dataset of (input, black-box prediction) pairs.
• Data augmentation: To improve surrogate fidelity, the input space can be sampled more densely around decision boundaries or uniformly across the feature distribution.

A global surrogate explains the entire model behavior across all input space, distinguishing it from local explanation methods.

• Global explanation: Answers 'How does the model generally make decisions?' using a single interpretable structure.
• Local explanation contrast: Methods like LIME explain a single prediction; a global surrogate provides one unified, approximate model.
• Trade-off: Global surrogates sacrifice local precision for a holistic overview, potentially missing nuanced, instance-specific decision boundaries.

The utility of a surrogate model is entirely dependent on its fidelity—how accurately it mimics the black-box model.

• High fidelity: The surrogate's decisions closely match the black-box, making it a reliable proxy for explanation.
• Low fidelity: The surrogate is a poor mimic, and any insights drawn from it are misleading artifacts of the surrogate's own limitations.
• Validation: Always measure and report the surrogate's performance against the black-box predictions on a held-out test set before interpreting its structure.

Global surrogates introduce a fundamental explanation gap because they model the black-box function, not the true data-generating process.

• Approximation error: The surrogate explains its own simplified logic, which may not perfectly capture the black-box's complex decision boundaries.
• Correlation vs. causation: The surrogate identifies correlational patterns in the black-box's predictions, not causal mechanisms.
• Feature masking: If the black-box uses complex interactions, a simple linear surrogate will fail to capture them, providing a dangerously incomplete picture.

Global surrogates are a primary tool for model auditing and regulatory compliance when the production model is a proprietary or uninterpretable ensemble.

• Regulatory documentation: A decision tree surrogate can be submitted to regulators as a transparent, auditable description of a deep neural network's logic.
• Debugging: If the surrogate reveals the black-box is relying on a spurious feature (e.g., a background pixel), engineers can identify and fix the flaw.
• Stakeholder communication: Provides a simplified mental model for non-technical stakeholders to understand general model drivers.