How machine learning powers Facebook's News Feed ranking algorithm

recommendation-systemsrankingmultitask-learningpersonalizationfacebook

Abstraction: Multi-pass multitask neural ranking system for News Feed personalization at scale

Key points:

Three-pass ranking architecture: pass 0 selects ~500 candidates via lightweight model; pass 1 scores each with deep multitask neural nets; pass 2 adds contextual diversity rules
Multitask learning predicts multiple engagement signals simultaneously (likes, comments, shares); combined into a single score V_ijt via weighted linear combination
Weights determined by user surveys asking which interactions feel meaningful and worth their time—not raw engagement rate alone
"Unread bumping" re-surfaces previously ranked-but-unseen posts; "action bumping" resurfaces posts with new comment activity since last session
System ranks 1,000+ posts per user per day for 2+ billion users in real time using embeddings, neural networks, and personalization based on observational data
Authors note long-term value (not short-term clicks) is the optimization target; survey alignment helps operationalize this

Connections: Facebook · Meta · Recommendation Systems · Multitask Learning · Ranking Algorithms