Decision-Focused Federated Learning Under Heterogeneous Objectives and Constraints

Researchers Konstantinos Ziliaskopoulos and Alexander Vinel have introduced a novel Decision-Focused Federated Learning (DFFL) framework that addresses a critical challenge in distributed AI: how to collaborate when different participants have fundamentally different optimization goals and constraints. Their approach builds on the SPO+ (Smart Predict-then-Optimize) method but extends it to federated settings where clients cannot share raw data. The core innovation lies in mathematically separating two types of heterogeneity: objective shift (differences in cost vectors) and feasible-set shift (differences in constraint sets), then providing formal bounds on how these differences affect learning performance.

Through rigorous analysis, the researchers developed a practical decision rule: federation improves decision quality when the penalty from client heterogeneity is smaller than the statistical advantage gained from pooling data. Their experiments using FedAvg-style implementations on both polyhedral and strongly convex optimization problems revealed important insights. Federation proves remarkably robust for strongly convex problems, maintaining performance even with noticeable differences between clients' optimization tasks. However, performance degrades in polyhedral settings primarily due to constraint heterogeneity, especially for clients with larger datasets.

The work provides both theoretical foundations and practical guidance for implementing federated learning in real-world scenarios where participants have legitimate reasons for different objectives and constraints—from healthcare institutions with varying patient populations to financial firms with different risk tolerances. By quantifying when collaboration is beneficial versus harmful, this research moves federated learning from a one-size-fits-all approach to a nuanced tool that can be deployed strategically based on mathematical guarantees.