Explicit preference elicitation is the usual workaround

What it is

With no behavioural history, the shortest patch is to ask: pick three films, tick a few classes, rate a few tracks. Explicit preference elicitation substitutes a statement the user will give now for a click sequence that does not yet exist, so the ranker has a non-zero vector on the first screen.

It is the usual workaround because a statement is faster than waiting for dozens of accidental clicks. It is not a truer preference — statements and later behaviour often disagree.

Why it happens

Elicitation rewrites cold start from an estimation problem into a form problem. Chosen seeds enter neighbourhood or content matching, and the first screen is no longer pure global popularity. Seed quality tracks how well the questions discriminate: all head titles, and the seed still sits in the prior; all obscure items, and new users cannot judge, so the seed is noisy.

A statement is a self-narrative of the moment, pulled by social desirability, memory, and “the listener I want to be.” That is not the same distribution as later clicks. The value of elicitation is a computable starting point, not a final profile. Later behaviour must be allowed to overwrite the seeds, or the first week’s ticks become a permanent bias.

Studying it

On a new-user cohort, compare: no elicitation (popularity baseline), short elicitation (few high-discrimination seeds), long elicitation. Offline, agreement between the seed vector and a week of real interactions (overlap, rank correlation). Online, first-screen relevance judgments, completion of elicitation, week-one retention. Independent variables: number of seeds, whether items or categories are asked, whether skip is allowed. Dependent variables: statement–behaviour agreement, first-screen usefulness, fraction who reach the home screen.

Success is not “they finished the form.” High completion with clicks a week later unrelated to the seeds means you elicited a story, not a usable signal.

Where it stops holding

On privacy-sensitive or low-involvement products (one-shot tools, anonymous browsing), asking does not arise; use context or stay on popularity. Children struggle with the meta-task of “represent my taste,” so seeds are less reliable. In enterprise catalogues preference is often a role, not a taste; elicit duties, not likes. This entry only argues that asking is the standard patch for cold start. It does not treat asking so much that people leave — that is the next cost.

Applying it

• Seed on a few concrete items, not a long strip of abstract tags. “Which of these three is closer to you” turns into a vector more readily than “pick twenty favourite labels.”
• Seeds must be overwritable by later behaviour: first-week clicks should rewrite first-day ticks, and that rewrite should be visible in the profile.
• Check: new users who finish elicitation should have a first screen with a nameable difference from the popularity baseline. After a week, overlap seeds with real clicks. Very low overlap means the elicitation was ceremony.

Related

• Same group: L6.03.1 Recommendation quality is lowest when there is no history · L6.03.3 Too many elicitation questions cause dropout before first use
• Nearby: L6.05 Turning Personalization Off · L6.06 Inferred Preferences and Their Limits · L6.10 Turning Personalization Off and Resetting It
• Search terms: explicit preference elicitation · cold-start onboarding · seed preferences