Model predictions are explained by learning a second, usually more explainable model, as a proxy. Surrogate model-based approaches are model-agnostic and can be used to achieve either local or global explanations.