Short Video Recommendation & Ranking

Design a short-video recommendation or ranking system for Spotlight / Lens-style products, covering retrieval, ranking, short-term interest, cold start, delayed labels, and multimodal features. Requirements Design a recommendation system for short videos, lenses, or a TikTok-like feed. Functional requirements: Generate a personalized feed for each user with fresh short videos or lenses. Capture short-term interest from recent watches, skips, likes, shares, hides, and follows. Support retrieval and ranking when only a minority of videos have text descriptions. Handle cold start for new users and new items. Train and serve models with enough observability to debug ranking regressions. Scale / constraints: Assume a large catalog, high request rate, and strict latency budget for feed vending. Most useful labels are delayed or biased: watch time, completion, rewatch, hide, share, and downstream retention. Content is multimodal: video frames, audio, text captions, creator, engagement history, and safety / quality signals. Design decisions: Candidate generation: two-tower embedding retrieval, ANN index, graph / creator co-engagement, trending pools, and business-rule sources. Ranking: lightweight online ranker vs heavier batch features; multi-task objectives for watch time, quality, and negative feedback. Freshness and exploration: mix exploitation with controlled exploration for new creators and new items. Bias handling: popularity bias, position bias, delayed labels, and feedback loops. Notes A strong answer separates candidate generation, feature hydration, ranking, re-ranking / diversification, logging, offline training, and online experimentation. For multimodal retrieval, CLIP-style contrastive embeddings are a natural tool, but they are not sufficient alone: text-light videos still need visual/audio embeddings, creator priors, engagement priors, and exploration. For short-term interest, discuss a session feature store and recent-event sequence features. The ranker should know whether the user just watched several similar videos, skipped a topic, or interacted with a creator. For cold start, use content embeddings, creator history, regional trends, and exploration buckets until personalized labels arrive. Evaluation should include offline ranking metrics, online A/B tests, guardrail metrics, creator ecosystem health, freshness, diversity, and negative feedback. Delayed labels can be handled with delayed attribution windows, counterfactual / debiasing techniques, and separate real-time proxies. Preparation Practice a 10-minute architecture: request -> candidate sources -> feature store -> ranker -> diversification -> logging -> training. Prepare crisp answers for two-tower retrieval, contrastive loss, ANN search, popularity bias, and cold start. Build a metrics table: watch time, completion, likes, shares, hides, long-term retention, diversity, and safety guardrails. Be ready to explain how retrieval and ranking stay aligned when they optimize different objectives.