Cold Start and Exploration

Every recsys interview ends with “what about new users/items?” Have a layered answer.

New items (item cold start)

1. Content-based representation: the item tower uses text/image/audio encoders (title, description, thumbnail via CLIP-style encoders, creator features) so a brand-new item embeds meaningfully with zero interactions. Train with ID-embedding dropout so the model doesn’t lean entirely on the ID (Two-Tower Retrieval Networks).
2. Exploration traffic: reserve a small slice of impressions for fresh items to buy interaction data. Allocate with bandits:
   • Epsilon-greedy: with prob ε show an explore item. Simple, wasteful.
   • UCB: score = predicted value + uncertainty bonus → explores what you know least.
   • Thompson sampling: maintain a posterior over each item’s quality (e.g., Beta over CTR); sample from posteriors and rank by samples — exploration emerges from uncertainty automatically; converges to exploitation as posteriors sharpen. The interview-favorite.
3. Graduation: once an item has enough impressions, its learned ID embedding takes over.

New users (user cold start)

1. Context features available from second zero: device, locale, time, acquisition channel, age band.
2. Onboarding signals: picked interests; first session’s clicks are gold — use a sequence encoder that produces useful states from 3 events.
3. Popularity-by-segment priors; shrink personalization toward segment means until data accumulates (hierarchical/empirical-Bayes mindset).
4. Meta-learning (mention only): MAML-style “learn an initialization that adapts fast per user” — rarely productionized; fine to name and move past.

The exploration–exploitation framing (say this)

Cold start is not a bug to patch; it’s the exploration side of a bandit: short-term engagement loss purchased for information. The design questions are budget (what % of impressions), placement (dedicated slots so exploration doesn’t pollute top ranks), and measurement (log propensities → enables off-policy evaluation).

Cold start in account matching

A brand-new account has no behavior for the behavioral encoder — yet registration time is exactly when ban-evasion matching matters most. Answer: at t=0 rely on hard signals (device/IP/payment/email + registration telemetry), produce a provisional embedding from the first session, and re-score as events accumulate (sequential evidence accumulation — see AM-09 Enforcement, Serving, and Infra). The system design must explicitly include this signal-maturity timeline.