LinkedIn Skills — Data Mining & ML System Design

Design the pipeline behind LinkedIn's Skills graph: how skill phrases are extracted from resumes / job posts, normalized into a canonical taxonomy, ranked per profile, and continually updated as new skills emerge. The round is graded as much on data collection / labeling strategy as on modeling. Requirements Functional: Build / maintain the canonical skills taxonomy (Java, Kubernetes, LLM Fine-Tuning, …). For each profile, output a ranked list of skills with confidence scores. For each job posting, extract required and preferred skills. New skills must surface as the language evolves — "RAG", "MoE", "Vector DB" did not exist five years ago. Non-functional: Profile coverage: a billion+ profiles, incremental updates as users edit content. Job posting coverage: tens of millions live, refreshed daily. Skill-prediction precision matters more than recall — false skill claims hurt user trust. Notes The expected structure: Mining layer — parse resumes / posts, surface skill candidate tokens (NER + dictionary lookup + LLM-assisted candidate expansion). Normalization — map surface form to canonical entity in the taxonomy (e.g. "Pytorch", "PyTorch", "pytorch" → canonical id). Per-profile scoring — features = mention frequency, recency, endorsements, role-keyword co-occurrence; output a confidence score per skill. Taxonomy maintenance — emerging-skill detection (token frequency over time, embedding-space novelty). Active learning loop: send low-confidence candidates to annotators, retrain. The interviewer expects an explicit conversation on labeling strategy — programmatic weak labels from co-occurrence vs annotator-curated seeds vs LLM-generated labels with a confidence threshold. Reported as an open question with the interviewer steering toward whichever sub-component the candidate hand-waves; treat each sub-component as 5 minutes of detail. Preparation Walk the four-layer pipeline (mining → normalization → scoring → taxonomy maintenance) end-to-end without prompting. Pre-load the data-collection discussion: weak labels, annotator throughput numbers, LLM-judge calibration. Be ready to discuss how the per-profile feature vector is constructed and what model produces the final score (gradient-boosted tree on a few hundred features is the canonical answer). For the emerging-skill follow-up, propose an embedding-novelty detector — new tokens whose embedding lies far from any current taxonomy node and whose frequency is rising.