Recommendation quality is lowest when there is no history

What it is

Collaborative filtering and most personalised rankers need “what this person clicked before.” A brand-new account, private mode, a cleared history, a device not yet merged — individual signal is near zero. Recommendation cold start is that lower bound: with no personal history, the list is global popularity, editorial picks, or a context guess, not this person.

What is cold is the user’s history, not the model. The model can be hot and still cold to this person.

Why it happens

Individual preference is a vector estimated from interactions. With zero observations the estimate falls back to a population prior: head titles, currently promoted items, broad classes tied to context (hour, place, device). The prior is best for the average user and high-variance for the person who just arrived. If the interface still labels the rail “for you,” the weakest estimate is being sold as personalization.

Item cold start is the symmetric half: an item with no interactions cannot enter collaborative neighbourhoods. User-cold and item-cold stack — a new user meeting a new title has no anchor on either side. Quality is lowest at that intersection, not in the cell “new user, old head titles.”

Studying it

Slice users by history length: zero interactions, 1–5, 6–20, more. On each slice run the same offline protocol (time-split recall and a ranking metric such as nDCG) and the same online contrast. Independent variables: available history length, whether context features are allowed, whether a social or device graph is allowed. Dependent variables: ranking metrics, first-screen clicks, rate of judging the list “not for me.”

Do not count a popular baseline on new users as personalization success. Popularity will almost always beat a personalised model on the cold slice — that is missing signal, not a good model. Reports must be cut by history length, or hot users will average away new-user failure.

Where it stops holding

Content matching that treats “the sentence I am searching” as a query is not personalization; cold start hurts it less. Strong-context openings (a running playlist, a shop QR) have a prior that is already enough. Forced login that imports history from another product line relocates cold start rather than solving it. This entry only claims that individual relevance is worst with no history. It does not require a questionnaire to fill the gap, and it does not treat questionnaires that drive people away.

Applying it

• With zero history, do not title the rail “for you.” Use “popular now” or “editors are watching,” so the label matches the estimate’s source.
• Treat cold start as its own channel: popularity plus editorial plus explicit context, rather than feeding an empty vector into a model tuned on veterans.
• Check: open the first screen on a fresh account and note ten rows; open it on an account with three months of history. If the two lists are nearly identical and both say “for you,” cold start is being dressed as personalization.

Related

• Same group: L6.03.2 Explicit preference elicitation is the usual workaround · L6.03.3 Too many elicitation questions cause dropout before first use
• Nearby: L6.09 Feedback Loops and Preference Entrenchment · L6.05 Turning Personalization Off · L6.04 Feedback Loops
• Search terms: recommendation cold start · new-user problem · item cold start