Problem
A social feed gets rare spam or policy-violating posts that pass the first filters at publication time. If such a post gets likes and reshares, it spreads through member networks and reaches many people before anyone removes it. Teams that build the defense have few labeled viral examples, and the spam types keep changing. The cost shows up as views of spam and as member reports.
Example from the article: LinkedIn found that some posts evaded its content defenses and were shared widely, so it added models that look at how a post is spreading.
Solution in brief
The defense has two layers. Proactive models score a post as soon as it appears, using content and author features. Reactive models rescore posts that already collect engagement, using post, member and engagement-velocity features, and they predict whether the viral post is spam. Positive predictions lead to automatic action or human review.
Questions this case answers
- How to detect harmful items that spread fast through a network before they reach many users, when positive examples are scarce and the first-pass filters miss them
- How to predict the harmfulness of an item from the early behavior of the users who react to it
- How to stop spam posts from going viral when the standard filters miss them at publication time?
- How to predict from early engagement whether a fast-growing post is spam?
Tools and methods
- Proactive classifiers — deep neural networks, one set per spam category (for example hate speech) and one per content type (video, article), run every few hours
- Reactive classifier — Boosted Trees model plus heuristics, scores posts that gather engagement
- TensorFlow — training framework for the proactive networks
- Human review queue — receives posts the models flag but do not remove automatically
Google or a one-time email code. No password.