# Rank Newly Launched Ads Under Cold Start Design a ranking model dedicated to ads launched within the last seven days, where direct performance history is sparse. ### Constraints & Assumptions - Practice assumption: the ranker scores eligible ads after policy, budget, and targeting filters. - The objective balances user value, advertiser value, and marketplace constraints. - Feedback includes impressions, clicks, conversions, hides, spend, and delayed conversion labels. - A new ad may have creative, campaign, advertiser, targeting, bid, and landing-page features. - The seven-day boundary comes from the prompt; the transition to the mature ranker must be defined. ### Clarifying Questions to Ask - Which auction quantity is ranked: click probability, conversion value, or expected utility? - How much exploration and advertiser risk are acceptable? - Can data transfer from related campaigns, creatives, or advertisers? - How delayed, censored, or biased are conversion labels? ### Part 1: Objective, Labels, and Features Define the score and training examples. Explain how content, context, advertiser priors, and uncertainty enter the model without leaking future data. #### Hints - Sparse ad-level history does not imply that all useful evidence is absent. #### What This Part Should Cover - Marketplace-aligned target - Cold-start feature hierarchy - Leakage and bias controls ### Part 2: Model and Exploration Choose a model architecture, shrinkage or transfer strategy, calibration method, and a controlled way to learn about new ads. #### Hints - A point estimate and confidence about that estimate answer different questions. #### What This Part Should Cover - Generalization with sparse labels - Uncertainty-aware exploration - Budget and safety constraints ### Part 3: Serving and Evaluation Design feature freshness, online scoring, handoff after seven days, experiments, monitoring, and fallback behavior. #### Hints - Abrupt model boundaries can create score discontinuities. #### What This Part Should Cover - Versioned low-latency serving - Counterfactual-aware evaluation - Smooth transition and operational guardrails ### What a Strong Answer Covers - A precise utility objective and label policy - Hierarchical evidence for new ads - Safe exploration with calibrated uncertainty - Serving, handoff, delayed-label evaluation, and marketplace guardrails ### Follow-up Questions - How would you evaluate an exploration policy offline? - How do you prevent large advertisers from dominating the prior? - What if an ad receives no impressions during its first day? - How would creative similarity help and potentially harm the model?