Cold-start to personalized feed transition

WHAT THIS TESTS This evaluates handling of the cold-start problem: serving something good with zero history and transitioning smoothly to personalization as evidence accumulates.

A GOOD ANSWER COVERS Bootstrap: with no data, serve a popularity or trending baseline so the feed is immediately useful; optionally seed with interests the user declares during onboarding to warm-start. Signals: collect implicit behavior, clicks, dwell time, scroll past, completion, skips, and saves, plus any explicit signals, and contextual ones like time and device. Transition: rather than a hard switch, blend a personalized score with the popularity baseline using a confidence weight that rises as the user accumulates interactions, so early on popularity dominates and over time personalization takes over. Keep some exploration so the feed does not collapse onto a narrow niche, and continually update the user representation as new signals arrive.

COMMON WRONG ANSWERS A binary cutover to full personalization after one or two clicks, overfitting noise. Ignoring onboarding-declared interests that could warm-start. Only using explicit signals, which are sparse, while ignoring rich implicit ones. No exploration, so the feed narrows and the user never discovers new categories.

LIKELY FOLLOW-UPS Why weight dwell time differently from a click? How do you balance exploration against exploitation? How do declared interests warm-start the model?

ONE CONCRETE EXAMPLE A brand-new user sees a trending feed plus a few items matching an interest they picked at signup. As they read three articles fully and skip two, the system raises the personalization weight; by the next session the feed blends roughly equal parts personalized and popular, and after sustained use it is predominantly personalized while still injecting occasional exploratory items.