Abstract
This work tackles the challenges of data heterogeneity and communicationlimitations in decentralized federated learning. We focus on creating acollaboration graph that guides each client in selecting suitable collaboratorsfor training personalized models that leverage their local data effectively.Our approach addresses these issues through a novel, communication-efficientstrategy that enhances resource efficiency. Unlike traditional methods, ourformulation identifies collaborators at a granular level by consideringcombinatorial relations of clients, enhancing personalization while minimizingcommunication overhead. We achieve this through a bi-level optimizationframework that employs a constrained greedy algorithm, resulting in aresource-efficient collaboration graph for personalized learning. Extensiveevaluation against various baselines across diverse datasets demonstrates thesuperiority of our method, named DPFL. DPFL consistently outperforms otherapproaches, showcasing its effectiveness in handling real-world dataheterogeneity, minimizing communication overhead, enhancing resourceefficiency, and building personalized models in decentralized federatedlearning scenarios.