Awesome Search KG

Building a Large-Scale Recommendation System: People You May Know

2 min read

Building a Large-Scale Recommendation System: People You May Know

Source: https://www.linkedin.com/blog/engineering/recommendations/building-a-large-scale-recommendation-system-people-you-may-know Company: LinkedIn

Summary

How LinkedIn scaled its People You May Know (PYMK) recommender to score >1 billion candidates per member while keeping relevance high and latency low. The design is a textbook Multi-Stage Ranking funnel: each stage narrows the candidate set for the next, optimizing a different objective with a different speed/quality trade-off.

The Four Stages

L0 — Candidate Generation

Selects a few thousand candidates from billions. Goal is recall, not order — ensure relevant candidates survive. Metric: Recall@k (k ≈ 3,000–5,000). Multiple candidate-generation sources: graph-based (random walks, n-hop neighbors), embedding-based retrieval (EBR), and heuristics (e.g. new members nearby).

L1 — Light Ranker

Takes the few thousand candidates, calibrates the diverse CG sources onto a common objective, and reduces to a few hundred. Lightweight model (logistic regression or XGBoost). Calibration matters because L0 mixes graph- and similarity-based candidates that aren’t directly comparable. Metric: Recall@k (k ≈ 500–800).

L2 — Rich Ranker

Heavy deep neural networks predicting probability and value of distinct engagement events (invitation sent, invitation accepted). Consumes the most powerful member–candidate pair features. Metrics: AUC, Precision@k, and ECE (expected calibration error) — scores are reused downstream, so calibration is tracked.

Re-Ranker

Final stage of fairness re-rankers (protected attributes), diversity re-rankers, and anti-overrepresentation logic. Multiple L2 engagement-event models are linearly combined, with the combination weights estimated via Bayesian optimization — a multi-objective tuning problem.

Why Offline ≠ Online

Each stage has offline metrics (Recall@k, AUC/Precision@k, log-likelihood/diversity), but the team trusts A/B tests for system-level judgment. Offline rarely matches online due to presentation/position bias, deployment errors, and train/serve distribution skew.

Why It Matters

A concrete production instance of the multi-stage pattern where the ranking objective changes per stage — recall first, then calibrated relevance, then engagement value, then fairness/diversity — illustrating that “the ranking objective” is rarely singular in real systems. See Ranking Objectives and Retrieval Pipeline.

Companies

Graph View

Backlinks