Diversity costs short-term clicks, so it has to be an independent objective

What it is

Slide something less like past clicks into the list and short-term clicks usually fall — people more often tap the familiar. If diversity is only a side curve on a dashboard, the next optimisation pass will trade it for clicks. Diversity as independent objective means diversity sits on the same sheet to maximise as clicks (or completion, dwell), and short-term clicks are allowed to give way to it, rather than waiting for clicks to “also” produce diversity.

That diversity has to be designed into the structure is the previous group’s claim. What is nailed here: if the design does not appear as an objective, clicks will pick it back out of the structure.

Why it happens

Clicks are a reward that settles immediately. Diversity’s return (discovery, long-run retention, less aesthetic fatigue) settles later and noisier. Reinforcement or online optimisation whose main reward is clicks will systematically underpay diversity. A side metric has no gradient, only a retrospective chart, and a chart does not win the next training pass.

Write diversity as a constraint (floor on coverage, cap on similarity) or as a weighted term, and it has force against clicks. The weight is a product decision: how many short-term clicks to spend for coverage. No weight means that decision is silently given to clicks.

Studying it

A Pareto front: sweep the diversity weight on one system, plot short-term clicks against coverage (or intra-list diversity). Online, interleave a clicks-only policy with clicks plus a diversity term; report two-week clicks, four-week retention and coverage. Independent variables: weight of diversity in the loss, regulariser versus hard constraint. Dependent variables: short-term clicks, coverage, later retention.

Do not discuss “diversity is free” under a condition where clicks are held equal. Held equal usually means the diversity weight was zero. Draw the cost, then decide whether to pay it. Offline nDCG leans toward the logging policy and cannot by itself veto a diversity term.

Where it stops holding

While the user is finishing a closed task, spending clicks on diversity is interference; drop the weight until the task ends. When inventory or safety has already pinned the set, a diversity term has no seat to buy. This entry argues that diversity must be able to fight clicks inside the objective. It does not choose random versus structured, and it does not choose the semantic dimension diversity should sit on.

Applying it

• Put coverage or intra-list diversity on the optimisation sheet beside clicks, with a written budget for how far first-screen clicks may fall. Without a budget the weight will be tuned back to zero before ship.
• Read two charts every week: clicks, coverage. A clicks-only week will silently kill the diversity term.
• Check: set the diversity weight to zero and watch whether coverage slides in a few days while clicks rise. If it does, diversity was never independent — it merely had not yet been eaten.

Related

• Same group: L6.08.1 Narrowing is produced by the objective itself; no one has to intend it · L6.08.3 Users cannot evaluate content they never saw; absence produces no visible feedback · L6.08.4 Random injection is not structured diversity; it only adds noise · L6.08.5 Diversity must sit on dimensions the user cares about; cross-category mix does not fix a single viewpoint
• Nearby: L6.02 Filter Bubbles · L6.09 Feedback Loops and Preference Entrenchment · L6.04 Feedback Loops
• Search terms: diversity as independent objective · first-class diversity · click–diversity tradeoff