A Federated Model Personalisation Method Based on Sparsity Representation and Clustering

As federated learning (FL) becomes more extensively employed, it attracts an increasing number of scholars and practitioners. In contrast to traditional decentralized machine learning approaches that acquire users' raw data, FL gathers locally updated gradients, protecting their privacy. However, different users may have disparate data distributions, resulting in underperformance of the federated model. It is beneficial to adapt the federated model to various data distributions. Numerous personalisation approaches have been examined, but most of them are limited to a single device with minimal data, making them susceptible to bias and overfitting. In fact, the data distributions of certain users are similar, and these similarities can be leveraged to increase the efficacy of personalisation. In this research, we describe a sparsity-based clustering method, as well as a federated personalisation strategy based on it. Our method mitigates the impact of non-IID data and generates more accurate local models. The trials reveal that it outperforms several of its counterparts.