Surrogate Model
A simple, interpretable model that approximates a complex black-box model to explain its decisions.
Surrogate models explain black-box AI by training a simple, interpretable model on the predictions of the complex model.
Explanation
Surrogate models train an interpretable model (e.g., decision tree) on the predictions of the black-box model. LIME uses local surrogates, global surrogates approximate the entire model.
Marketing Relevance
Surrogate models are the basis of LIME and enable explainability without access to model internals.
Common Pitfalls
Surrogate only approximates – cannot perfectly represent the black-box model. Fidelity must be measured.
Origin & History
The concept comes from simulation optimization in the 1970s. Ribeiro et al. used local surrogate models in LIME in 2016. Global surrogates became popular in XAI research as an alternative to SHAP.
Comparisons & Differences
Surrogate Model vs. SHAP
SHAP computes exact feature contributions with Shapley values; surrogate models approximate behavior with a simple model.
Surrogate Model vs. Distillation
Knowledge distillation trains a model for prediction; surrogate models are trained for explanation.