Design a Personalized Feed and Compare Ranking Objectives

Design a personalized professional-network news feed, similar in purpose to LinkedIn's feed. Explain the recommendation system from data and features through ranking and serving. Compare pointwise, pairwise, and listwise learning-to-rank methods in detail. ### Part 1 — Data and features Describe the candidate content, training examples, labels, and features needed to personalize the feed. #### What This Part Should Cover - Member, content, relationship, and request-context information that is available at prediction time. - Exposure logging and the distinction between an unshown item and a shown item with no interaction. - Delayed outcomes, leakage prevention, and new-member or new-content behavior. ### Part 2 — Ranking methods Compare pointwise, pairwise, and listwise approaches. Explain their training units, objective functions, and practical tradeoffs for a feed, then choose a starting approach. #### What This Part Should Cover - Scoring individual examples, learning preferences between items, and optimizing a slate or list. - Grouping examples by the relevant member/request context. - The relationship between the training loss, position-sensitive ranking quality, and product objectives. ### Part 3 — Serving and evaluation Describe how the recommendation system produces a feed within an agreed serving budget and how you would evaluate improvements. #### What This Part Should Cover - Candidate retrieval, filtering, feature lookup, ranking, and final feed assembly. - Consistent features and model versions, with a fallback when a dependency is slow. - Offline evaluation followed by an online experiment with meaningful guardrails. ### What a Strong Answer Covers - A coherent path from exposure and feedback data to a served personalized list. - A ranking-method choice justified by data quality, objective, cost, and iteration needs. - Explicit handling of feedback bias and missing history without claiming knowledge of a company's internal system. ### Follow-up Questions - Why can a model trained on clicked and unclicked impressions inherit position bias? - When would a better offline ranking metric fail to improve the experience of a new member?