Personalized federated meta-learning based on dynamic clustering

Due to the issues of poor convergence on highly heterogeneous data and the lack of personalized solutions faced by traditional Federated Learning (FL), Personalized Federated Learning (PFL) has emerged as a current research hotspot and trend. PFL aims to provide customized models for each client by considering the data characteristics and requirements of different clients (or devices), thereby enhancing model accuracy and practicality. Meta-learning has become one of the key technologies for realizing PFL due to its rapid adaptation capabilities. By learning how to effectively adjust and optimize learning algorithms from historical tasks, meta-learning enables PFL systems to more quickly adapt to new clients and datasets. However, in scenarios with heterogeneous data, a single global model approach in meta-learning may struggle to achieve good performance. In this work, we propose a dynamic clustering-based personalized federated meta-learning framework (C2FML), which clusters clients based on the low-dimensional representations of their data participating in training during each iteration. Subsequently, within each cluster, the model updates obtained through meta-learning using local data from clients belonging to that cluster are aggregated to obtain a personalized model for that cluster, which is then distributed to the respective clients. This way, each client obtains a dedicated personalized model upon completion of training. Our research ensures that each group consists of clients with similar data distributions, which facilitates model learning in multitask scenarios and accelerates model convergence. By doing so, the proposed framework can be applied in highly heterogeneous environments, and we evaluate the effectiveness of our framework on public datasets. Experimental results demonstrate that, compared to state-of-the-art federated meta-learning, our proposed framework achieves improved model accuracy.