Early accidental clicks get amplified into durable profile features

What it is

A new user taps a curiosity headline, a child grabs the phone and opens a cartoon, a thumb hits an ad — that one positive enters the profile, and weeks of “for you” orbit it. Accidental-click entrenchment is a single, early, intention-unclear positive written as a stable preference feature instead of dropped as noise.

The loop amplifying early rank advantage is a Matthew effect on the item side. Here it is on the user vector: one mis-tap becomes “who this person is.”

Why it happens

At cold start, positives are few and each one weighs a lot. The model has no channel for “this was a mis-tap,” so a click is preference. Later display orbits that point, and true preference has not yet produced enough positives to overwrite it. Noise is written into the most stable place at the moment the sample is thinnest.

Unclear intent also includes: tapping for someone else, buying a gift, a one-off task (fix a pipe). In a human narrative those are episodes; in the counter they are class mass. Without decay, without down-weighting a single low-completion, the episode lives as a lifetime tag. When the user later says “that is not me,” the system still hears week-one’s count.

Studying it

On new users, inject one controlled “not like me” click (a specified item under informed consent) against no injection. Weeks later, measure that class’s weight in the profile, its exposure share, and whether the user can recognise it as “that mis-tap.” Independent variables: at which interaction the injection happens, whether completion is extremely short, whether a decay window exists. Dependent variables: residue of the class in the later vector, success at clearing it with “not me.”

Report item-side Matthew (does that title stay on the first screen) separately from profile-side freeze (does the class become a feature). If the title vanishes and the class remains, the feature has already frozen.

Where it stops holding

A first choice the user explicitly offered as a seed (three films in an elicitation form) is not an accident; it is a statement. Early clicks with high completion and repeated returns should not be treated as noise. Shared accounts write several people’s mis-taps into one vector, which makes freeze worse, but that is an account-model problem. This entry is “intention-unclear early positives freeze as features.” It does not treat later performance, and it does not treat the immediate cost of exploration.

Applying it

• Decay positives while history is below a threshold: down-weight or mark as suspect those with low completion, a single occurrence, or disagreement with later behaviour.
• Give “that was not me” a path that hits the feature, not only the item, especially in week one.
• Check: on a test account, mis-tap a class far from later real use on day one, and look at that class’s exposure on day seven. If it still takes a serious share of the first screen, and clearing the item does not clear the class, freeze has happened.

Related

• Same group: L6.09.1 The system only observes items it showed; unshown items never receive positive feedback · L6.09.3 Users adjust behaviour to get the recommendations they want, so behavioural data lose representativeness · L6.09.4 Breaking the loop needs active exploration; the cost is immediate and the benefit is delayed · L6.09.5 Offline evaluation on historical logs systematically favours the policy that produced those logs
• Nearby: L6.04 Feedback Loops · L6.03 Cold Start · L6.10 Turning Personalization Off and Resetting It
• Search terms: accidental-click entrenchment · noisy first clicks · profile freeze