Bandits for Online Calibration: Application to Content Moderation at Meta

Speaker: Deeksha Sinha, Research Scientist at Meta

What tech lies behind the social media giants’ attempts to keep content ”within the rules”? At Meta, we have both hand-crafted and learned risk models to flag questionable content, for humans to review. To operationalize these, we aggregate the different models to give a single ranking score, calibrating them to prioritize more reliable risk models. But violation trends change over time, affecting which risk models are most reliable; risk models change; and novel models are introduced. To continuously update the system in response to such trends, we use a contextual bandit. Our approach increases Meta’s top-line metric for measuring the effectiveness of its content moderation strategy by 13%.