The paper presents a new algorithm for Federated Learning (FL) that identifies beneficial collaborations among clients. The algorithm assigns adaptive aggregation weights to clients based on the conduciveness of their data distributions to a specific learning objective. The study demonstrates that collaborations guided by this algorithm outperform traditional FL approaches. The findings highlight the importance of careful client selection and suggest potential improvements for FL implementations in the future.
Publication date: 7 Feb 2024
Project Page: https://arxiv.org/abs/2402.05050v1
Paper: https://arxiv.org/pdf/2402.05050
