Preventing Abuse Using Unsupervised Learning | Fighting Abuse @Scale 2019

Detection of abusive activity on a large social network is an adversarial challenge: attack patterns evolve quickly and ground truth labels are imperfect. These characteristics limit supervised learning approaches, but they can be overcome with unsupervised methods. In this talk from Fighting Abuse @Scale 2019, I cover how we used isolation forests to detect abusive automation at LinkedIn scale, why label-free outlier detection holds up against adapting adversaries, and results from production. The distributed Scala/Spark implementation I built was open sourced as linkedin/isolation-forest. Code: https://github.com/linkedin/isolation... Write-up: https://jverbus.github.io/2019/08/13/... More of my work: https://jverbus.github.io/ Originally recorded and published by the @Scale conference: https://atscaleconference.com/videos/...