Anchors

Anchors are an interpretability method that explains a model's prediction by finding a decision rule (an "anchor") that is sufficient to guarantee the prediction — if the anchor conditions are met, the prediction is (almost) always the same, regardless of other feature values.

How Anchors Work

• Rule Format: IF (feature_1 = value_1) AND (feature_2 = value_2) THEN prediction = class_A (with precision ≥ τ).
• Precision: The fraction of instances matching the anchor that have the same prediction (e.g., τ = 95%).
• Search: Use beam search with perturbation-based coverage estimation to find the shortest sufficient anchor.
• Coverage: The fraction of all instances where the anchor applies — wider coverage = more general rule.

Why It Matters

• Sufficient Explanations: Unlike LIME/SHAP (which show feature importance), anchors give sufficient conditions for the prediction.
• Actionable: An anchor rule is directly actionable — "as long as these conditions hold, the prediction won't change."
• Model-Agnostic: Works with any classifier — just needs black-box access.

Anchors are sufficient explanation rules — finding the simplest set of conditions that lock in a prediction regardless of other features.