LinkedIn Feed Ranking Architecture

LinkedIn's candidate generation uses a two-tower neural network: one tower encodes the member (profile, interests, network) into an embedding, and the other tower encodes each post (content, author, engagement signals) into an embedding. The member's embedding is used for approximate nearest-neighbor search against millions of post embeddings to retrieve the top-K (~1,000) candidates. This separation is key: post embeddings are precomputed offline, so retrieval at serving time is just a vector similarity search — completing in milliseconds even over millions of candidates. The two-tower architecture sacrifices fine-grained member-post interaction features for massive speed.

The second-pass model scores the top few hundred candidates using gradient-boosted decision trees (GBDTs) and deep neural networks consuming 1,000+ features. Features span multiple categories: member features (seniority, industry, activity level), post features (content type, author authority, freshness), interaction features (has the member engaged with this author before, content-topic affinity), and context features (time of day, device, session depth). GBDTs handle sparse categorical features well, while neural networks capture complex feature interactions. The model predicts multiple engagement signals — probability of like, comment, share, and long dwell — combined via a multi-objective function.

LinkedIn's ranking models consume features from two pipelines. The offline pipeline computes batch features daily using Spark: a member's industry affinity scores, author authority metrics, content topic distributions, and network graph features. These are materialized into a key-value feature store (Venice). The online pipeline computes real-time features per request: current session behavior, trending topics, recent engagement signals, and freshness decay. The feature store serves both pipelines with sub-5ms p99 latency. Ensuring consistency between training features (computed offline) and serving features (computed in real-time) is a major engineering challenge that LinkedIn addresses via feature logging and monitoring.

After scoring, LinkedIn applies diversity injection to prevent feeds from becoming monotonous. Without intervention, the ranking model would fill feeds with the highest-scoring content type (often viral engagement-bait posts). Diversity rules enforce variety across dimensions: content type (articles, images, videos, job posts), author diversity (don't show 5 posts from the same person), topic diversity (mix industry news with career advice), and network distance (blend first-degree connection posts with suggested content). These rules are applied via a post-processing diversification algorithm that re-ranks the scored list subject to diversity constraints.

LinkedIn actively demotes content that optimizes for engagement at the expense of quality — 'engagement bait' like polls designed to game the algorithm. A quality scoring model evaluates content integrity: is this post informative or clickbait? Is the author an authority on this topic? Does it contribute to professional discourse? Posts that receive engagement primarily from people outside the author's professional network are flagged as potentially viral-but-low-quality. LinkedIn also applies velocity controls — rapidly viral posts are held for human review before being amplified further. This reflects LinkedIn's explicit product decision to prioritize professional value over engagement metrics.