Model-Agnostic

Explainability methods that can be applied to any ML model.

The only look at input-output dependencies and have no way of looking at internal structures of the models.

So these methods are Post-Hoc methods. Often times we use these two definitions interchangeably.

Pros

• good seperation from ML model
• flexibility of model and method choice
• decouple interpretability layer and ml layer

Cons

• low Fidelity
• often only local