Federated Learning With Only Positive Labels by Exploring Label Correlations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training across decentralized data sources while preserving user privacy.
• Multilabel classification (MLC) in FL settings faces challenges, particularly with imbalanced positive-only data per client, leading to poor performance.
• Existing methods using server-side regularizers or frequent private embedding exchange have limitations in addressing label correlations and communication costs.

Purpose of the Study:

• To propose a novel federated learning method that effectively addresses the challenges of multilabel classification.
• To leverage label correlations within the FL framework to enhance model performance.
• To improve the safety and reduce communication overhead in federated MLC.

Main Methods:

• Federated Averaging with Label Correlations (FedALC) estimates and utilizes label correlations in class embedding learning.
• A variant of FedALC proposes learning fixed class embeddings per client to minimize server-client communication.
• The method focuses on improving model training by exploiting relationships between different label pairs.

Main Results:

• FedALC significantly outperforms existing federated MLC methods across multiple datasets.
• The proposed variant enhances data safety and reduces communication overhead by exchanging class embeddings only once.
• The approach effectively utilizes label correlations for improved classification accuracy.

Conclusions:

• FedALC offers a robust and efficient solution for multilabel classification in federated learning settings.
• The method successfully tackles the limitations of previous approaches by incorporating label correlations and optimizing communication.
• This work provides a valuable advancement for privacy-preserving collaborative machine learning.