New Federated Learning Methods Address Performative Prediction Challenges

Federated bilevel optimization is a common approach for nested learning problems across multiple distributed clients, used in areas like hyperparameter tuning and meta-learning. However, existing models often assume static client data distributions, which is frequently violated in practice due to "performativity"—where deployed decisions influence client behavior and data collection, causing decision-dependent distribution shifts. This paper introduces the concept of federated bilevel performative prediction, where both the upper-level and lower-level objectives are evaluated under these dynamic, client-specific, decision-dependent distributions. The researchers formalize the "federated bilevel performatively stable" (FBPS) point and establish conditions for its existence and uniqueness. To compute this FBPS solution, two novel federated methods are developed: FBi-RRM, which offers linear convergence under specific conditions, and FBi-SGD, a communication-efficient stochastic method. Both methods come with convergence guarantees, particularly under diminishing step sizes and sufficiently small sensitivities. Experiments confirm the predicted stability thresholds and show improved meta-generalization compared to non-performative baselines, even in complex nonconvex neural network settings.