FedDSS: A data-similarity approach for client selection in horizontal federated learning

Area of Science:

• Machine Learning
• Distributed Systems
• Healthcare Informatics

Background:

• Federated learning (FL) enables collaborative model training without sharing sensitive data, addressing healthcare data challenges.
• Non-independent and identically distributed (non-i.i.d.) data across clients causes model divergence and performance degradation in FL.
• Existing FL methods struggle with heterogeneous data distributions, impacting model accuracy.

Purpose of the Study:

• To introduce FedDSS (Federated Data Similarity Selection), a novel FL framework designed to mitigate performance issues caused by non-i.i.d. data.
• To enhance model convergence and predictive accuracy in FL settings by employing a data-similarity client selection strategy.
• To ensure client data privacy is maintained throughout the federated learning process.

Main Methods:

• FedDSS utilizes a statistical data similarity metric, an N-similar-neighbor network, and a network-based client selection strategy.
• Performance was evaluated against FedAvg using pediatric sepsis datasets (PICD, MIMICIII) in both i.i.d. and non-i.i.d. scenarios.
• Key metrics included average loss, true positive rate (TPR), and selection fairness (entropy).

Main Results:

• FedDSS demonstrated faster convergence and higher True Positive Rates (TPR) compared to FedAvg in both i.i.d. and non-i.i.d. settings.
• On PICD, FedDSS achieved higher TPR earlier than FedAvg. On MIMICIII, FedDSS showed significant loss reduction and faster TPR achievement.
• FedDSS exhibited reasonable selection fairness across different data distributions.

Conclusions:

• FedDSS effectively improves federated learning performance by enabling faster convergence and achieving desired TPR with fewer communication rounds.
• The proposed data-similarity approach enhances sepsis prediction accuracy within the FL framework.
• FedDSS offers a privacy-preserving solution to the challenges posed by non-i.i.d. data in healthcare FL applications.