LinkedIn Learning / Job Recommendation Ranking

Design a recommender that surfaces LinkedIn Learning courses (or jobs) based on the user's profile, recently applied jobs, and behavioural history. Standard candidate-generation + ranking architecture; the LinkedIn-specific twist is heavy feature engineering on the job-posting side and explicit explainability for the recommended item. Requirements Functional: Given a user, return a ranked list of courses (or jobs) the user is likely to engage with. Cold-start handling for new users / new items. Explainability — why is this course recommended? (e.g. "skills you applied for last week"). Non-functional: Sub-200 ms candidate generation + ranking budget. Recall over the full catalog; precision over the top-10 served. Online learning loop closes within hours, not days. Notes Two-stage architecture: Candidate generation — retrieve ~1000 candidates per request. Two-tower retrieval (user tower + item tower) trained on co-engagement; ANN index over item embeddings. Ranking — gradient-boosted tree (or a deep ranker) over ~hundreds of features. Score the ~1000 candidates, take top-K. Feature engineering on the job / course side is the recurring deep-dive: parse the job posting, extract required skills, level, location, salary band; cross with user profile features. Calibration of the ranker's output probability matters when downstream business logic (promoted course budgets, position eligibility) consumes it — discuss isotonic regression or Platt scaling. Multi-objective ranking is a frequent follow-up: blend engagement, business value, fairness. Common framing is a linear blend with weights tuned offline and validated A/B online. Preparation Walk the two-stage architecture end-to-end in < 15 minutes. Drill the feature-engineering discussion — for both user and item sides, list ~10 features each with a one-line motivation. Pre-load the A/B testing discussion: counterfactual evaluation, treatment-control split, novelty effects, business metric vs proxy metric. For the explainability follow-up, propose feature-attribution (top-3 features by contribution) rather than per-prediction LIME / SHAP — production explanations are usually rule-based attributions of the top driver.