Federated Bilevel Performative Prediction
Researchers have introduced Federated Bilevel Performative Prediction (FBi-PP) to address challenges in federated learning where deployed decisions can alter client behavior and data distributions. This new framework accounts for decision-dependent distribution shifts at both the upper and lower levels of optimization. Two methods, FBi-RRM and FBi-SGD, are proposed to compute the performatively stable solution, with FBi-RRM offering linear convergence and FBi-SGD providing communication efficiency. Experiments using CNNs for classification and strategic regression tasks demonstrated the effectiveness of these methods in improving meta-generalization compared to non-performative approaches. AI
IMPACT Introduces novel methods for handling complex distribution shifts in federated learning, potentially improving model robustness and generalization in real-world distributed systems.