Multistage Ranking for LinkedIn Feed
Thursday, September 19, 2019 — 4:30PM - 5:30PM
LinkedIn’s personalized feed is made up of news, articles, jobs, and activities from the member’s professional network. Whenever a user visits their feed, a large heterogeneous set of feed updates need to be retrieved and ranked in a scalable manner. In order to accomplish this goal, we utilize a multistage ranking system, where the earlier stages focus on candidate generation over homogenous feed updates and the later stages focus on ranking and blending of heterogeneous feed update candidates incorporating business objectives such as freshness and diversity. In this paper, we have given an overview of this system and our approach in designing and evaluating the machine learning models involved at different stages.
Madhulekha Arunmozhi, LinkedIn