Global Surrogate

Unlike local surrogates (e.g., LIME) that explain a single prediction, a global surrogate is trained on a large dataset of input-output pairs from the black-box model to mimic its behavior everywhere. The goal is to learn a function g that approximates the black-box f such that g(x) ≈ f(x) for all x in the input space. This provides a single, unified explanation of the model's overall logic.

• Training Data: Generated by querying the black-box model with inputs, often using the original training set or a uniform sample from the feature space.
• Fidelity Metric: Measured by the R² score or accuracy of the surrogate on held-out black-box predictions.

The surrogate must be a transparent-by-design model whose decision logic can be directly inspected by a human. The choice of student model determines the form of the explanation.

• Decision Tree Surrogate: Produces a flowchart of if-then rules. Visual depth and node count indicate complexity.
• Linear Proxy Model: Yields feature weights that represent global importance and direction of influence.
• Explainable Boosting Machine (EBM): A glass-box model that provides per-feature shape functions showing how each variable contributes to predictions.
• Rule Lists: Algorithms like SIRUS extract a compact set of stable if-then rules.

The core tension in global surrogacy is between fidelity (how accurately the surrogate matches the black-box) and interpretability (how simple the surrogate is to understand). A highly complex black-box may require a surrogate so complex that it ceases to be interpretable.

• High Fidelity, Low Interpretability: A deep decision tree with hundreds of leaves may match the black-box well but is unreadable.
• Low Fidelity, High Interpretability: A shallow tree or linear model is easy to understand but may miss critical non-linear interactions.
• Evaluation: Always report fidelity metrics alongside the surrogate. A surrogate with 70% fidelity may be useless for auditing, even if it is perfectly readable.

Global surrogate methods are model-agnostic, meaning they treat the original black-box model as an opaque oracle. No access to gradients, internal weights, or architecture details is required. The surrogate only needs the ability to query the model with inputs and receive predictions.

• Applicability: Works on any model type—deep neural networks, gradient-boosted trees, ensembles, or even external APIs.
• Separation of Concerns: The explanation method is completely decoupled from the model's training process. This is a post-hoc explanation technique applied after the black-box is fully trained and frozen.

Global surrogates suffer from a fundamental flaw: they explain the surrogate model, not the black-box. If fidelity is imperfect, the explanation is misleading. Additionally, surrogates can be highly unstable.

• Sampling Sensitivity: The surrogate's structure can change dramatically based on the specific inputs used to query the black-box for training data.
• Extrapolation Blindness: The surrogate only learns from the black-box's outputs on the sampled data. It may fail catastrophically in regions of the input space that were not sampled.
• Complexity Mismatch: A simple surrogate cannot capture a fundamentally complex decision boundary. The explanation will be a gross oversimplification.

A specialized form of global surrogacy is rule extraction, where the student model is constrained to produce a set of symbolic if-then rules. This is often achieved through rule-regularized distillation.

• Process: Train a neural network with a penalty on weight complexity, then extract logical rules from the simplified network.
• Pedagogical vs. Decompositional: Pedagogical approaches treat the model as an oracle (standard surrogacy). Decompositional approaches inspect individual neuron activations.
• Output: A human-readable rule set like "IF age > 60 AND income < 50k THEN high risk" that approximates the black-box logic.

In lending, a deep neural network may predict default risk with high accuracy, but regulators like the CFPB require adverse action reasons. A global surrogate, often a decision tree or Explainable Boosting Machine (EBM), is trained on the black-box model's input-output pairs. This surrogate provides a complete, interpretable rule set (e.g., 'IF debt-to-income ratio > 0.43 AND number of recent inquiries > 3 THEN high risk') that approximates the original model's decision boundary across the entire applicant population, enabling compliance with the Equal Credit Opportunity Act.

A convolutional neural network (CNN) analyzing chest X-rays may detect pneumonia with superhuman accuracy, but clinicians require transparent reasoning to trust a diagnosis. A global surrogate, such as a decision tree surrogate or rule list, is trained to mimic the CNN's classifications across the entire dataset. The extracted rules (e.g., 'IF opacity in lower left lobe AND fever present THEN positive diagnosis') provide a high-level, human-readable approximation of the model's logic. This allows for clinical validation against established medical knowledge before deployment in a hospital PACS system.

A complex gradient-boosted tree ensemble predicts equipment failure from thousands of IoT sensor streams. To audit this model for safety-critical machinery, engineers train a global linear proxy model or shallow decision tree on the ensemble's predictions. The surrogate reveals the dominant global failure signatures, such as 'vibration amplitude in bearing 3' and 'temperature delta across the casing'. This allows engineers to validate that the black-box model's overall logic aligns with known mechanical physics and failure modes, rather than relying on spurious correlations.

Actuaries use complex models to price policies, but must ensure they do not use illegally discriminatory variables like race or gender, even indirectly. A global surrogate, often a Generalized Additive Model (GAM) or Explainable Boosting Machine, is trained to replicate the pricing model's output. The GAM's shape functions visually display the exact contribution of each permitted feature (e.g., age, driving history) across its entire range. This allows compliance officers to verify that no prohibited feature has a systematic, global influence on the final premium, satisfying model risk management (MRM) requirements.

In drug discovery, a graph neural network (GNN) may predict molecular toxicity with high accuracy, but its internal reasoning is opaque. Scientists train a global rule extraction algorithm on the GNN's predictions to distill its knowledge into a set of symbolic chemical rules (e.g., 'IF molecule contains a phenol group AND a halogen in the para position THEN toxic'). These extracted rules, representing the model's complete learned logic, serve as a hypothesis generator, suggesting new structure-activity relationships (SARs) for medicinal chemists to investigate and validate experimentally.

An end-to-end deep reinforcement learning policy controls a vehicle's steering and acceleration. To build a safety case for regulators, a global surrogate model, such as a CART decision tree, is trained on millions of state-action pairs from the policy. This provides a complete, approximate map of the vehicle's behavioral logic: 'IF distance to lead vehicle < 10m AND speed > 30 kph THEN decelerate'. Safety engineers can then exhaustively audit this surrogate tree to identify any globally unsafe branches or logic contradictions before the system is allowed on public roads.