Anchors: Rule-Based Explanations

Methods like LIME explain a single prediction by fitting a simple linear model that approximates the black-box model's behavior nearby — but a linear approximation can still be misleading if the true decision boundary curves sharply close to that point. Anchors take a different approach: instead of a weighted linear formula, an anchor is a short, human-readable if-then rule (like "if income > $80k AND credit history is clean") that is verified, by testing many nearby synthetic inputs, to hold the model's prediction fixed with high probability whenever the rule is satisfied.

Because an anchor is checked empirically rather than assumed from a local linear fit, it comes with a precision guarantee: if the rule says the prediction stays the same with 95% precision, that number was measured by actually perturbing inputs and checking, not estimated from a model. The tradeoff is coverage — a very precise, tightly-scoped anchor rule might only apply to a narrow slice of inputs, while a broader rule risks lower precision.

Worked example

A loan-approval model denies an applicant. An anchor explanation might report the rule "debt-to-income ratio > 45% AND at least one late payment in the last 12 months" with 96% precision — meaning that among many synthetic applicants who satisfy exactly that rule but vary everything else, 96% were also denied by the model. This gives the applicant (and the auditor) a concrete, testable condition driving the decision, rather than a set of feature weights whose combined effect is harder to interpret directly.