Which approach uses a simpler, interpretable model to approximate the behavior of a complex black-box model locally?

Which approach uses a simpler, interpretable model to approximate the behavior of a complex black-box model locally?

• A. Shapley Values
• B. Surrogate Models
• C. Retrieval Augmented Generation
• D. Counterfactual Explanations

Answer: (B)

Surrogate models provide a simple, interpretable stand-in that locally mimics a complex black-box model. The idea is to focus on a small neighborhood around the instance you're examining: generate nearby samples, get the black-box’s predictions for those samples, and then fit a lightweight, easy-to-interpret model (like a linear model or small decision tree) to those predictions, giving more weight to points closer to the original instance. This local surrogate captures how the black-box behaves in that region, so you can read off which features push the prediction up or down. The explanation is grounded in fidelity to the local decision surface, which is exactly what you need when you want an understandable summary of the model’s behavior for a specific case. Shapley values, while powerful for attributing a prediction to individual features, are not a single, simple surrogate model of the local behavior; they provide feature contributions rather than a stand-in model. Retrieval augmented generation is about bringing in external information to improve outputs, not about explaining a model’s local decision process. Counterfactual explanations describe what changes to inputs would flip the outcome, which is a different form of explanation than building a local, interpretable proxy.

Surrogate models provide a simple, interpretable stand-in that locally mimics a complex black-box model. The idea is to focus on a small neighborhood around the instance you're examining: generate nearby samples, get the black-box’s predictions for those samples, and then fit a lightweight, easy-to-interpret model (like a linear model or small decision tree) to those predictions, giving more weight to points closer to the original instance. This local surrogate captures how the black-box behaves in that region, so you can read off which features push the prediction up or down. The explanation is grounded in fidelity to the local decision surface, which is exactly what you need when you want an understandable summary of the model’s behavior for a specific case.

Shapley values, while powerful for attributing a prediction to individual features, are not a single, simple surrogate model of the local behavior; they provide feature contributions rather than a stand-in model. Retrieval augmented generation is about bringing in external information to improve outputs, not about explaining a model’s local decision process. Counterfactual explanations describe what changes to inputs would flip the outcome, which is a different form of explanation than building a local, interpretable proxy.