DA-PFL: Dynamic Affinity Aggregation in Personalized Federated Learning Under Class Imbalance

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Personalized federated learning (PFL) aims to create tailored models for each client.
• Current PFL methods often aggregate clients with similar data distributions.
• This similarity-based aggregation can worsen the class imbalance problem in datasets.

Purpose of the Study:

• To propose a novel dynamic affinity-based PFL (DA-PFL) model.
• To alleviate the class imbalance issue inherent in federated learning.
• To enhance the performance of personalized learning models.

Main Methods:

• Developed a complementary affinity metric to guide client aggregation.
• Implemented a dynamic aggregation strategy adjusting client selection each round.
• Evaluated the DA-PFL model on four real-world datasets.

Main Results:

• The DA-PFL model significantly improved client accuracy.
• Effectively reduced the negative impact of class imbalance during federated learning.
• Outperformed existing state-of-the-art comparison methods.

Conclusions:

• The proposed DA-PFL model offers an effective solution for class imbalance in PFL.
• Dynamic aggregation based on affinity enhances personalized model performance.
• DA-PFL demonstrates superior accuracy across diverse datasets.