Candidate generation

For Feed posts, a Microsoft patent describes building the shortlist through first-pass rankers, systems that find and initially score possible recommendations.1 The patent lists rankers for articles, jobs, news and hashtag activities, plus followfeed, a component that selects possible feed updates; this is not a statement of how LinkedIn works today.1 LinkedIn Engineering's May 2026 write-up describes FishDB as the layer that stores Feed data and performs initial selection of possible Feed posts.2 FishDB keeps organized lookup records in memory to help it find possible posts quickly.2

Another way to build a shortlist is to compare embeddings, lists of numbers that describe members and items so the closest matches can be found.35 LinkedIn Engineering's August 2026 post says it used a large language model adapted through further training to produce these lists and match members to creators to follow.3 It built one process that prepares suggestions in advance and another that retrieves them when needed, to meet different requirements for response speed, freshness and scale.3 The advance process handles large batches, such as nightly runs, and compares each viewer's list of numbers with every creator available for recommendation—fewer than 10 million—to find the closest matches.3 A Microsoft patent from July 2026 describes a large language model that creates numerical descriptions for content and users in a shared comparison space, a network with separate user and content parts that adjusts those descriptions using user engagement, and matching by similarity to select an audience for content.5

For job matching, a September 2026 LinkedIn research paper describes how numerical matching can also balance different goals when building a shortlist.4 The paper says common retrieval models combine relevance and business goals such as engagement and revenue into a single fixed numerical representation, so a trained model cannot adapt to shifting priorities while running without retraining.4 Its method, Embedding Subspace Partitioning, divides each numerical representation into parts for different goals, then combines the similarity scores for those parts using weights that can be adjusted while the system runs.4 The authors report that in LinkedIn's job matching platform, the method enabled changes to retrieval priorities while running and significant improvements in key business metrics.4