Reasons must faithfully reflect the ranking basis

What it is

The ranker actually used “completion in the last three days.” The interface writes “because you like jazz.” Even if the user does like jazz, the sentence does not report the basis. A faithful recommendation reason is one whose cited cause was a real contributor to this item’s current position, not a user fact that happens to be true.

Happening to be true is biography. Constituting the basis is an explanation.

Why it happens

People invert the reason to rebuild the system. If the reason says “jazz” and a pop hit appears next, the system is read as lying or broken. A false internal model is worse than no model: forecasts now run on a rule that does not exist. Acceptance follows that false rule — the person is accepting a system that is not there.

Contribution in a ranker splits into feature weights, neighbours, context. A reason that quotes only the nicest factor is selective citation. Users cannot correct a model from the omitted terms, so “faithful” has to cover factors that moved rank, not every true sentence in the profile.

Studying it

The fidelity check is a correspondence: take the features or neighbours with largest contribution in the ranker, and code whether the interface text cites them. Independent variables: copy drawn from real contribution versus rewritten as a true-but-zero-contribution user fact. Dependent variables: whether people can later name the real basis, judgments of “did the system lie,” accuracy of using the reason to forecast the next item.

Do not treat “the reason sounded credible” as fidelity. Plausible copy can be wholly unfaithful. Post-hoc generative narration is a different problem. The question here is only whether the sentence points at a real contribution.

Where it stops holding

Contribution in deep models is itself unstable (ablation and Shapley values can name different lead factors for the same row). “True” can only be as sharp as the model family allows, not unique causation. Signals that cannot be disclosed (auction, quota) leave every reason incomplete; mark “other unlisted factors” rather than inventing a substitute story. This entry does not treat persuasion-to-open, and it does not treat boilerplate phrasing.

Applying it

• Write the reason from the largest contributor in the ranking log. Do not pick a nicer sentence from the profile.
• If the lead factor cannot be said (ads, inventory), write “promoted” or “stock first.” Do not fill the gap with a taste claim.
• Check: zero the named feature in the model; this item’s rank should drop. If rank does not move, the reason is false. Then ask users “why is it here?” — answers should land on that feature, not on any true line from the profile.

Related

• Same group: L6.01.1 Reasons raise predictability and acceptance · L6.01.3 A generic reason is no reason
• Nearby: L6.07 Presenting Recommendation Reasons · L5.08 Counterfactual Explanations · L5.02 Local and Global Explanations
• Search terms: faithful recommendation reason · explanation fidelity · feature attribution