Please note: This master’s thesis presentation will take place in DC 2310.
Zehao Zhang, Master’s candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Pascal Poupart
Asymmetric clustering represents a critical yet under-explored challenge in Clustered Federated Learning (CFL). Existing methods often compromise data utilization or model accuracy by either separating devices with different data quality into distinct clusters or merging all devices into a single cluster. The need for asymmetric clustering arises in practical scenarios where not all devices contribute equally due to varying data quality or quantity. For example, in healthcare, devices at a research hospital might generate high-quality medical imaging data compared to a small clinic. Asymmetric clustering allows high-quality data sources to enhance the learning of models on devices with lower-quality data without the need for reciprocity, which is crucial in such imbalanced environments.
We introduce a novel federated learning technique that enables selective contributions from some devices to others’ model training without requiring equal give-and-take. Crucially, our approach excels in the Federated Continual Learning (FCL) setting by addressing temporal heterogeneity and concept drift through its ensemble features.
Through detailed empirical evaluations, we validate that our approach not only efficiently generates high- quality asymmetric clustering but also significantly enhances performance in continual learning settings. This adaptability makes it highly suitable for real-world applications where data distributions are not static but evolve over time.