Connected Content Retriever: Dense Graph Edge Features Powering Pre-Ranking at LinkedIn

In the LinkedIn Feed, content from a member's network (connections and follows) accounts for over 70% of impressions and engagement, so the pre-ranking layer must forward the best possible few hundred candidates to ranking. LinkedIn's professional knowledge graph provides engagement signals across first- and second-degree networks, including stranger-viral posts that a 1st-degree connection reacted to, commented on, or reshared without authoring. This fan-out pushes the candidate index above one billion, while selecting activities from the viewer's network narrows it to roughly tens of thousands that must be scored within a 120 ms p99 latency budget.

The paper presents Connected Content Retriever (CC Retriever), a pre-ranking system that scores these candidates using a full deep ranking model on GPUs at low latency. Its core is a sorted-search GPU primitive that joins dense graph affinity features between viewer and author with document-level features stored on the GPU at runtime, performing this join in 5–10 ms.

Shifting to GPU-served scoring enabled a 50x scale-up of the ranking model's parameters and delivered a +2.5% lift in content time spent on the LinkedIn Feed in online experiments, significantly higher than typical gains observed in LinkedIn Feed experiments. The work describes the feature set drawn from LinkedIn's economic graph and the scoring model architecture, with particular emphasis on the online system that scales the stack.