Problem
Yelp added video uploads to reviews in April 2023, and the number of uploaded videos rose sharply. Videos with nudity, graphic violence or extremist imagery are rare, but they must be hidden before users see them. Any classifier will flag some legitimate videos, and removing them frustrates users. Human moderators can restore them, yet manual review is slow and does not scale. Videos are also much larger than reviews or photos, so running them through a neural network takes much more time, while the classification has to be near real time.
Example from the article: a video uploaded with a review is checked in parallel with ingestion and is either discarded, hidden for review, or left published.
Solution in brief
A three-stage pipeline: a similarity-hash match against previously removed videos, a deep learning multi-label classifier on sampled frames, and human review of videos whose scores pass tuned thresholds. Frame sampling and blocking suspicious uploaders cut inference time. The frame classifier is built on the existing photo moderation model.
Questions this case answers
- How to screen large media uploads for policy violations in near real time when running a neural network on every frame is too slow and human review capacity is limited
- How to keep false positives of an automatic content filter low enough that human reviewers are not overloaded
- How to moderate user-uploaded videos in near real time without sending every video through a slow neural network?
- How to keep the false positive rate of an automatic moderation model low when human reviewers restore the wrongly flagged content?
Tools and methods
- Similarity hash matching service — discards videos that match earlier removals
- Deep learning multi-label classifier — scores sampled frames by content category
- Score thresholds — set to keep false positives minimal while still catching violations
- Frame sampling — passes only selected frames to the model to cut inference time
- Human review (User Operations team) — removes true violations and restores false positives
Google or a one-time email code. No password.