Towards Fairness in Personalized Ads Using Impression Variance Aware Reinforcement Learning

Demographic representation gaps in ads are measurable and fixable: The paper demonstrates that ad impressions are often distributed unevenly across demographic groups, and Meta developed metrics to quantify these "fairness gaps" for protected characteristics like gender and estimated race. Their Variance Reduction System (VRS) successfully reduced these disparities by re-ranking which ads users see.

Fairness means matching eligible audiences: Rather than showing ads equally to all users, VRS aims to align the demographics of people who actually see an ad with the demographics of people who are eligible to see it based on advertiser targeting criteria. This approach respects advertiser intent while reducing unequal outcomes.

Privacy-preserving fairness is achievable at scale: The framework operates in a privacy-preserving manner while handling multiple protected characteristics simultaneously, demonstrating that fairness considerations don't require compromising user privacy—an important consideration for responsible AI deployment.

Real-world results show meaningful impact: VRS has been deployed to all US users for housing ads, where it produced "significant improvement" in fairness metrics. This represents the first large-scale deployment of a fairness framework addressing multiple demographic groups in online advertising.

Fairness in ads requires ongoing technical innovation: The paper shows that achieving equitable outcomes in personalized advertising systems requires specialized algorithmic approaches (reinforcement learning-based re-ranking), suggesting fairness isn't automatic and demands deliberate engineering effort.