Understanding anchor explanations
In Chapter 6, Local Model-Agnostic Interpretation Methods, we learned that LIME trains a local surrogate model (specifically a weighted sparse linear model) on a perturbed version of your dataset in the neighborhood of your instance of interest. The result is that you approximate a local decision boundary that can help you interpret the model's prediction for it.
Like LIME, anchors are also derived from a model-agnostic perturbation-based strategy. However, they are not about the decision boundary but the decision region. Anchors are also known as scoped rules because they list some decision rules that apply to your instance and its perturbed neighborhood. This neighborhood is also known as the perturbation space. An important detail is to what extent the rules apply to it, known as precision.
Imagine the neighborhood around your instance. You would expect the points to have more similar predictions the closer you get to your instance, right...
