MLSD — Reels / Short Video Recommendation

Design the Reels short-video recommender. Four-stage retrieval → pre-rank → rank → re-rank scaffold is table stakes; differentiation is in features, metrics, and A/B-test design. Requirements Functional: serve a personalized feed of short videos to each user; optimize for engagement (watch-time, completion, share). Four-stage scaffold (expected to be present from minute 1): Retrieval / candidate generation: Two-Tower model or ANN over embeddings; collaborative + content signals. Pre-rank: lightweight DCN-V2 / DLRM to cut from ~10k candidates to ~hundreds. Rank: heavy model — DCN-V2 + MMoE for multi-task (CTR, complete-rate, comment-rate, share-rate). Re-rank: business rules + diversity (MMR) + exploration (epsilon-greedy / Thompson sampling). Features: user (demographics, recent watch history), video (creator, topic, age, embeddings), interaction (cross features), context (time, device), social (GraphSAGE on friend graph). Metrics: online A/B (watch-time, daily-active, retention); offline (AUC, recall@k, NDCG); calibration. Bias mitigation: position bias via IPW; training/serving skew monitoring; feedback-loop dampening. Cold start: content-based fallback for new users/videos; explore bandits. Notes Reports note interviewers now care more about metric design, online + offline eval, and feature engineering than about novel architecture. Lead with metrics, not models. One candidate's session focused entirely on A/B test design and feature collection — be ready for that emphasis. Preparation Memorize the 4-stage diagram so you can draw and annotate it in 3 minutes. Prep a metric tree: north-star (DAU retention) → proxy (watch-time, completion) → model objective (CTR, MMoE heads) → guardrails (sessions, complaints). Drill A/B-test design: hypothesis, randomization unit, MDE / sample-size, segment cuts, novelty effect. Practice the position-bias + IPW pitch.