## Question Design a recommendation system for a short-video feed. The system should generate a personalized ordered feed, react to recent viewing behavior, handle new users and new videos, and balance engagement with content diversity and exploration. ### Constraints & Assumptions - The catalog and user population are large enough that scoring every video per request is impossible. - Available signals include impressions, watch duration, completion, likes, comments, follows, skips, creator relationships, and content metadata or embeddings. - Recent behavior should influence the next recommendations quickly. - The feed must avoid excessive repetition from one creator or content type. - The product cares about durable consumption and retention, not click-through rate alone. ### Clarifying Questions to Ask - What is the primary product objective, and which guardrail metrics must not regress? - What latency and freshness targets apply to feed generation and feature updates? - Are there policy, safety, age, geography, or inventory constraints before ranking? - How much exploration traffic may be allocated to uncertain items? - Which interactions count as positive, negative, or ambiguous feedback? ```hint Separate scale-reduction stages Use multiple candidate sources, then progressively more expensive ranking models, followed by rule-aware re-ranking. ``` ```hint Freeze features at event time Offline training examples must contain only features that would have been available when the impression was served. ``` ### What a Strong Answer Covers - Multi-source retrieval, deduplication, lightweight pre-ranking, multi-objective ranking, and diversity-aware re-ranking. - User, item, context, sequence, and creator features with offline and real-time feature paths. - New-user and new-video exploration strategies with bounded risk. - Point-in-time-correct training data, negative sampling, delayed labels, and leakage prevention. - Offline evaluation followed by randomized online experiments on consumption and retention, with safety and quality guardrails. - Serving latency, caching, model rollout, fallbacks, observability, and feedback-loop risks. ### Follow-up Questions 1. How would you choose or learn weights across completion, likes, comments, and follows? 2. How would you make a skip influence the very next feed request? 3. How would you prevent popular creators from crowding out exploration? 4. What signals would reveal training-serving skew or feature leakage?