Learning pathsA
Production designs

Design video recommendations

Separate candidate generation from ranking.

Understand the problem

Ingest impressions and interactions with reliable timestamps. Generate candidates from embeddings, subscriptions, and popular items. Rank a bounded set, apply policy filters, and log exposure for evaluation.

Make it concrete

Training on clicks alone ignores unseen items; exposure logs help distinguish no interest from no opportunity.

Trade-offs and pitfalls

Optimize engagement alongside diversity, safety, and long-term user satisfaction.

Check your understanding

Handle new users, new items, stale features, and a ranking outage.

Practice this topic

Your study notes