With only positive signals the system cannot tell "not interested" from "never saw it"

What it is

The log has clicks and no “saw it, do not want it.” An unclicked recommendation may have been not scrolled to, scrolled to and not parsed, parsed and uninteresting, or meant to be tapped and missed. Positive-only feedback ambiguity means that without an explicit negative channel, a non-click cannot be translated as rejection. The system can see likes that happened, not dislikes that happened.

Exposure bias says unshown items get no positive. What is nailed here: even among the shown, without a negative channel, rejection stays unobservable.

Why it happens

Positives are sparse and high-confidence; non-clicks are numerous and low-confidence. Treat non-click as negative, and “never saw” is trained as “hate”; the model learns to show less in positions that need a scroll, and the ambiguity is written into policy. Do not treat non-click as negative, and the model lacks rejection samples and can only expand the positive neighbourhood — the bubble sits more firmly. Both roads lack a “I saw this, no.”

An explicit negative channel splits “seen and rejected” out of the mass of non-clicks. It does not remove exposure bias (unshown items still have no label), but it makes rejection among the shown learnable. Without it, negatives are either wrongly manufactured or absent.

Studying it

Contrast: positives only, non-click as negative, plus an explicit “not interested.” Use probe slots so a slice of items is known to have been seen. Dependent variables: calibration of scoring “seen and skipped” as negative, rate of wrongly punishing unseen items, whether later exposure dares positions that need a scroll. Independent variables: whether a negative channel exists, whether an operational trace of “seen” is required (dwell, impression duration).

Impression duration can be weak evidence; it is not a negative channel — people can stare blankly. Report whether ambiguity falls when a clear reject act exists. Do not invent percentage points on a dataset.

Where it stops holding

Forced-choice tasks (this question must be answered) have no legal semantics for “not interested.” Unclicks on pure search results are closer to “not this intent”; the negative channel should be labelled “this result does not match,” not “I dislike me.” This entry only treats that positive signals cannot carry rejection in recommendation. It does not specify the scope of a negative, and it does not treat emotional tapping.

Applying it

• Give shown items a clear reject act. Keep non-click as missing, not as negative.
• Store a reject only when there is evidence of exposure (entered the viewport and dwelt). Do not learn a flash of a row as hate.
• Check: the same item, one group explicitly rejects, one group never scrolled to it. If the model gives both similar negative scores, the ambiguity is still there. The reject group should be clearly more negative; the unscrolled group should sit near missing.

Related

• Same group: L6.13.2 Negative feedback must state its scope: this item or this class · L6.13.3 Negative feedback must produce an immediately visible change, or users will judge it useless and stop · L6.13.4 Users treat negative feedback as emotional expression; its semantics are imprecise · L6.13.5 Over-responding to a single negative can wipe out a whole class of content
• Nearby: L6.09 Feedback Loops and Preference Entrenchment · L6.08 Filter Bubbles and Diversity · L6.07 Presenting Recommendation Reasons
• Search terms: positive-only feedback ambiguity · explicit negative feedback · seen but rejected