Over-responding to a single negative can wipe out a whole class of content

What it is

A badly made crime show is marked disliked, and the whole crime class leaves home, including the title the user follows every week. Over-generalized negative feedback is one reject written into a unit that is too large, so a class, an author cluster, or a topic is wiped by mistake, and the cost of repair exceeds that one discomfort.

Scope has to be declared. What is nailed here is default and strength: even if the chosen scope is “this item,” if the gradient hits the class, the result is still a wipe.

Why it happens

Negatives are sparse; models like to squeeze the most parameter update out of one reject. Nearest neighbours, topic vectors, tag trees all propagate the update along similarity edges. A reject of a bad title becomes a penalty on the features “crime,” “night,” “this studio.” What the user sees is the class carried off. An affective pulse meeting that propagation hurts more: it was never a stable proposition, yet it is written at the maximum radius of one.

Over-width also shuts discovery: once the class is empty, probe slots and diversity quotas cannot fill that dimension, and silent missing coverage follows. If undo restores only the row and not the class-level weights, people think they have undone it and the class is still gone.

Studying it

On accounts known to like a class, reject one low-quality item in it, default scope item. Measure: exposure change for the rest of the class, liked items collaterally hit, how long until the user notices the wipe. Independent variables: whether the update stops at item, at author, or propagates along the tag tree; whether a count threshold exists (how many same-class rejects before promoting to class). Dependent variables: drop in class-level exposure, collateral count, whether the class returns after undo.

Do not take “they tapped dislike so the class should shrink” as success. Success is the scope the user later draws, not the radius the model propagated on its own.

Where it stops holding

Safety rejects should propagate; that is a different channel. Several rejects in a row of high-quality items in the same class reasonably promote to class, but the promotion should be visible and reversible. When items have no fine class, only a coarse one, any class-level update is already wide and should stop at item or author. This entry is over-wide propagation. It does not restate affective semantics, and it does not treat immediate visibility as already a guard against over-width — what vanishes at once may be the whole class.

Applying it

• One reject updates the item by default (plus author, if the user chose author). Class-level updates want a count threshold or a second confirm, and a cap on how far they walk along the tag tree.
• Undo must roll class-level weights back with the row, not only insert the row into the list.
• Check: on a test account that likes crime shows, reject one bad title and watch crime exposure for a week. If the followed series leaves with it, propagation is too wide. Add a threshold, reject once more, and the followed series should still be there.

Related

• Same group: L6.13.1 With only positive signals the system cannot tell "not interested" from "never saw it" · 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
• Nearby: L6.08 Filter Bubbles and Diversity · L6.09 Feedback Loops and Preference Entrenchment · L6.10 Turning Personalization Off and Resetting It
• Search terms: over-generalized negative feedback · one-tap wipe · negative transfer