Problem
LinkedIn's human reviewers get hundreds of thousands of items per week in the content review queue. Handled first in, first out (FIFO), the queue wastes reviewer time on items that turn out to be non-violative (cleared). Violating content also stays visible longer when it arrives behind cleared items. The older prioritization used a few static buckets with different service-level agreements (SLAs) and could not react to new evidence about an item.
Illustrative example: a violating post arrives in the queue behind a long run of cleared items and is not reviewed until they are done.
Solution in brief
Every item entering the queue is scored by XGBoost models for the probability of a policy violation. The score is refreshed on each new member report. Items likely to be clear are deprioritized, likely violations move up, and a part of the queue is decided automatically at a very high precision bar.
Questions this case answers
- How to order a human review queue by model-predicted risk so that scarce reviewer capacity goes to the items most likely to need action
- How to automate a share of review decisions at a fixed very high precision while the pattern of violations keeps changing
- How to prioritize a human content review queue by the predicted probability of a policy violation?
- How to cut the time until violating content is caught when reviewers work through a backlog?
Tools and methods
- XGBoost — models that predict the probability that an item is violative or clear.
- Threshold selection at high precision — model chosen by recall at extremely high precision (R@P), which also sets the auto-decision zone.
- Dynamic review prioritization — queue order driven by continuously updated scores plus a few static parameters.
- Human in the loop — reviewers keep the severe and ambiguous items.
Google or a one-time email code. No password.