COPF: An Online Framework for Deployment-Stable Counterfactual Fairness in Evolving Graphs

Researchers introduce COPF, a framework for monitoring and controlling fairness in online link recommendation systems on evolving graphs. The system addresses the challenge that recommendation algorithms are performative—they change user behavior and create feedback loops that make traditional fairness estimates unreliable after deployment.

COPF tackles a fundamental problem in recommendation systems: fairness metrics computed from historical data become invalid once a new recommendation policy is deployed, because the system's choices influence which outcomes are observed. Traditional fairness auditing assumes static data distributions, but online link recommendation systems are inherently dynamic. The framework introduces exposure-based fairness metrics using counterfactual reasoning, making them measurable through explicit exploration and propensity logging. This addresses a critical gap where deployed systems can experience significant fairness drift despite appearing fair during development. The research contributes graph-aware doubly robust estimators and residual outcome indistinguishability testing, providing both theoretical guarantees and practical auditing mechanisms. Experiments demonstrate that COPF can reduce worst-case fairness violations while maintaining reasonable ranking utility. The work represents meaningful progress in algorithmic fairness for systems where recommendations actively shape data generation. For practitioners deploying recommendation systems in hiring, lending, social networks, or other high-stakes domains, deployment-stable fairness monitoring addresses a genuine operational challenge. The explicit connection between fairness theory and online learning provides a foundation for responsible system updates. Going forward, the key questions involve scaling these methods to production systems with millions of users and understanding how to balance fairness guarantees against user experience metrics.

• →COPF provides deployment-stable fairness monitoring for recommendation systems where recommendations influence future outcomes
• →The framework uses counterfactual reasoning and propensity logging to make fairness gaps measurable in evolving systems
• →Theoretical guarantees connect residual outcome indistinguishability to bounds on exposure-based group disparities
• →Experimental results show fairness improvements with modest ranking utility costs on real and synthetic datasets
• →Open-source implementation enables practitioners to audit and control fairness in deployed link recommendation systems