Beyond Independent Manipulation: Individual Fairness-aware Strategic Classification with Peer Imitation

Researchers introduce Individual Fairness-aware Strategic Classification (IFSC), a framework addressing how agents manipulate features when machine learning models prioritize individual fairness. Unlike existing approaches assuming independent agent behavior, IFSC models peer-driven manipulation where agents imitate nearby positively-decided peers, using robust learning to handle uncertainty in peer observability.

This research addresses a critical gap in strategic classification theory by recognizing that individual fairness constraints fundamentally alter how rational agents behave. Traditional strategic classification assumes agents independently optimize their feature manipulations; however, individual fairness requirements—which mandate similar treatment for similar individuals—create interdependencies where each agent's optimal strategy depends on neighbors' outcomes. This mismatch between classical models and fairness-aware settings has practical implications for deployed ML systems in lending, hiring, and admissions.

The IFSC framework models a realistic manipulation dynamic: agents observe which similar peers receive favorable decisions and imitate their characteristics to improve their own outcomes. This peer-imitation model better captures actual behavior in high-stakes domains than independent manipulation assumptions. The robust learning approach acknowledging uncertainty about which peers' decisions agents observe addresses a key practical challenge in fairness implementation.

For AI practitioners and ML governance, this research highlights an overlooked externality of individual fairness policies. While individual fairness sounds desirable, its transparency regarding who gets accepted creates cascading imitation effects that can distort the population distribution and undermine the fairness guarantees it intended to provide. The empirical validation on synthetic and real datasets suggests practitioners must account for these strategic responses when designing fair classification systems.

The work influences ongoing debates about fairness definitions in high-stakes machine learning. As organizations implement individual fairness constraints in response to regulatory pressure, understanding how these constraints reshape agent behavior becomes essential for predicting real-world outcomes and designing more robust fair classifiers.

• →Individual fairness constraints create interdependent agent manipulation through peer imitation, unlike independent behavior assumed in classical strategic classification.
• →IFSC models agents imitating positively-decided similar peers, more accurately capturing real-world manipulation in fairness-aware systems.
• →Robust learning with stochastic perturbations addresses uncertainty about which peer outcomes agents observe during manipulation.
• →Individual fairness policies can create unintended cascading effects that distort population distributions and undermine fairness guarantees.
• →Practitioners must account for strategic responses when implementing individual fairness to ensure classifier robustness and actual fairness outcomes.