Surrogate Models

Surrogate models treat the original fraud detection system as a complete black box, requiring no access to its internal weights, gradients, or architecture. They learn exclusively from the input-output pairs of the parent model.

• Universal Applicability: Works identically on gradient-boosted trees, deep neural networks, or proprietary vendor APIs.
• Non-Intrusive: No modification to the production model is required, preserving existing validation and compliance certifications.
• Training Data: The surrogate is trained on a synthetic or sampled dataset labeled by the black-box model's predictions, not the original ground truth.

The defining constraint of a surrogate model is that its internal structure must be natively transparent. Common choices include shallow decision trees, logistic regression, or Explainable Boosting Machines (EBMs).

• Decision Trees: Provide a direct, rule-based path to a fraud score, easily converted to reason codes for adverse action notices.
• Linear Models: Offer clear feature weights, showing the exact marginal contribution of transaction amount or time-of-day to the risk score.
• Glass-Box Requirement: The surrogate's value is entirely dependent on its own interpretability; a complex neural network surrogate defeats the purpose.

A surrogate model aims for global fidelity—accurately mimicking the black-box model's predictions across the entire input space—rather than explaining a single transaction. This is a distinct goal from local methods like LIME or SHAP.

• Fidelity Metric: Measured by the R-squared value between the black-box and surrogate predictions on a held-out test set.
• Trade-off: Higher fidelity often requires a more complex surrogate, which reduces interpretability. A 95% fidelity shallow tree is often preferred over a 99% fidelity deep tree.
• Global Insight: Reveals the overall decision strategy, such as "the model primarily flags transactions over $10,000 from high-risk jurisdictions," rather than explaining a single alert.

The surrogate model performs a form of knowledge distillation, compressing the complex, non-linear feature interactions learned by the black-box into a simplified, human-readable feature set.

• Feature Selection: Often trained on a reduced set of the most important features identified through preliminary permutation importance analysis.
• Interaction Transparency: Glass-box surrogates like EBMs explicitly show pairwise feature interactions, revealing if the black-box learned that "high amount AND foreign IP" is a multiplicative risk factor.
• Discretization: Continuous features may be binned to create categorical rules, such as "Transaction Amount > $9,000" as a binary flag, making the logic auditable by compliance teams.

For financial fraud and anti-money laundering (AML) use cases, surrogate models provide a stable, deterministic explanation layer that satisfies model risk management (MRM) and regulatory requirements.

• Deterministic Output: Unlike some local explanation methods that can have stochastic components, a trained surrogate produces the same explanation for the same input every time.
• Audit Trail Integration: The surrogate's decision path can be serialized and stored as part of the algorithmic audit trail for every flagged transaction.
• Model Governance: Surrogates are themselves models and must be versioned, validated, and monitored for fidelity drift as the parent model is retrained, fitting into existing MLOps governance frameworks.

Surrogate models introduce a layer of approximation that can be misleading if not rigorously validated. They explain the mimic, not the original.

• Fidelity Gaps: A surrogate may achieve high global fidelity but fail to capture critical decision boundaries in sparse regions of the feature space where novel fraud patterns emerge.
• Misattribution Risk: Correlated features in the training data can cause the surrogate to attribute importance to a proxy feature rather than the true causal driver used by the black-box.
• Complexity Ceiling: If the black-box model's decision function is highly discontinuous or relies on complex feature engineering, no simple surrogate will achieve acceptable fidelity, rendering the approach ineffective.