Problem
LinkedIn recommends people to connect with or to catch up with ("People You May Know", PYMK). The member base is more than one billion, so scoring and sorting every member for every viewer is prohibitively expensive. A first stage must cut this pool to a short list that the ranker can score.
Example from the article: for a viewer Alice, her 2-hop and 3-hop neighbors in the connection graph become candidates, even though she is not yet connected to them.
Solution in brief
Candidates are pooled from three sources: graph walks (N-hop neighbors and personalized PageRank), embedding similarity from a two-tower model, and simple rules. The ranker gets each candidate's score and source as features. The embedding model is trained on impressed candidates as positives and on a mix of three negative types.
Questions this case answers
- How to retrieve a short list of relevant candidates for each user from a pool too large to score exhaustively, when the items are nodes in a graph?
- How to train an embedding retrieval model on logged impressions without biasing it against popular or rarely active entities?
- How to pick a short list of relevant people to recommend from a pool of a billion members?
- How to find candidates that are far away in the social graph for users with few connections?
Tools and methods
- N-hop neighbors — shallow deterministic graph walk, 2 and 3 hops, ranked by common connections
- Personalized PageRank (PPR) — deep probabilistic walk with teleportation limited to the viewer's neighborhood
- Two-tower model for embedding based retrieval (EBR) — member embeddings for similarity search
- In-batch, random and nearest-neighbor negatives — negative sampling for the two-tower loss
- Approximate nearest-neighbor search — retrieval of the closest candidate embeddings
- Heuristic filters — recent feed interactions, notification clicks, profile searches
Google or a one-time email code. No password.