We present a model-agnostic federated learning method for networks of heterogeneous data and models. The network structure reflects similarities between the (statistics of the) local datasets and, in turn, their associated local (personal) models. Our method is an instance of empirical risk minimization, with a regularization term derived from the network structure of the data. In particular, we require well-connected local models, which form clusters, to yield similar predictions on shared public, unlabelled dataset(s). The proposed method allows for a wide range of local models. The only restriction is that these local models must allow for efficient implementation of regularized empirical risk minimization (training). For many models, such implementations are readily available in high-level programming libraries, including scikit-learn, Keras, and PyTorch.
View on arXiv@article{abdurakhmanova2025_2302.04363,
title={ Towards Model-Agnostic Federated Learning over Networks },
author={ S. Abdurakhmanova and Y. SarcheshmehPour and A. Jung },
journal={arXiv preprint arXiv:2302.04363},
year={ 2025 }
}