LinkedIn Professional Network  ·  View 30 of 30  ·  7 · Assurance

Abuse Detection and Content Moderation

Rules, models, member reports and human reviewers, all writing one decision log that can be audited.

Editable source SVG draw.io All views
Signal Detect Decide Act Audit Fake accounts Signup + login events Account risk model cluster detection Risk threshold Challenge / restrict Decision log Content PostCreated Classifiers spam · hate · links Confidence routing Hide · label · limit Decision log Member reports Report button Dedupe + priority Human review T&S queue Action + notify Appeal record Fraudulent jobs JobCreated Job fraud model Hold for review Close + ban poster Decision log low confidence Abuse Detection and Content Moderation Queue / topic Application we own Decision point Security / platform Data store Person or role synchronous Rules, models, reports and people all write the same decision log, so every action can be appealed. v 1.0 · owner Trust & Safety Engineering · date 2026-09

Decisions

  • Four lanes, one decision log: rules, models, reports and reviewers record decisions in the same way, so any action can be audited and appealed
  • Confident classifications act automatically; uncertain ones go to human review
  • CASAL is LinkedIn's anti-abuse application layer (2023). The open-source isolation-forest library flags anomalous accounts (2019)

Signals

  • A machine-learned risk score at registration, plus detection of account clusters (LinkedIn, 2018)
  • Content classifiers run nearline on PostCreated and again when a post is reported

Risks

  • Adversaries adapt, so models retrain on reviewer decisions
  • Over-enforcement hurts legitimate creators, so appeals have an SLA
◀  Privacy and Authorization All views
Solution Architecture v1.0 · LinkedIn's own stack in LinkedIn-run colos · Enterprise Architecture · 2026-0930-abuse-and-moderation