Rebuilding LinkedIn's Follows Recommendations with LLM-Based Semantic Retrieval and Ranking

Haohua Wan, Da Xu, Cindy Liang, Nishant Satya Lakshmikanth, Jeevan Gregory Sequeira24 Aug 2026

Problem

LinkedIn recommends creators to follow on the MyNetwork tab and Home Feed. The old multi-stage system used heuristic candidate generation and popularity-driven ranking. It worked for obvious, popular creators but missed good matches that were not popular. New members, who have almost no activity, got the weakest recommendations.

Example from the article: a robotics engineer would value a lesser-known researcher writing about embodied AI, but a popularity-based system rarely shows that creator.

Solution in brief

Write each member and creator profile as a short natural-language prompt and encode it with a fine-tuned LLM embedding model, so both sides live in one vector space. Use that space for retrieval (offline batch and online real-time) next to the old generators, and reuse the same embeddings, compressed to 64 or 128 dimensions, as features in the ranking model.

Questions this case answers
  • How to retrieve and rank items for users with no interaction history when popularity-based generators surface only obvious, already-popular items
  • How to reuse one learned text-embedding space for both candidate retrieval and the ranking model so that candidates from different sources are scored consistently
  • How to recommend creators to follow when popularity-based candidate generation misses relevant but less popular creators?
  • How to give relevant follow recommendations to new members who have almost no activity history?

Tools and methods

  • Narrative prompting — turns semi-structured profile fields into a natural-language description before encoding
  • Bi-encoder with InfoNCE loss — fine-tunes a shared encoder so followed creators sit close to the viewer
  • LoRA, mixed precision, gradient checkpointing — make fine-tuning affordable at production scale
  • Ray — scales batch embedding inference across a GPU cluster
  • FAISS (IndexFlatIP) — exact nearest-neighbor search over creator embeddings
  • Supervised projection layers — compress embeddings for the ranking model
Continue reading — it's free

Google or a one-time email code. No password.