1: \begin{abstract}

2: % The abstract must be limited to one paragraph.

3: In federated learning, model personalization can be a very effective strategy to deal with  heterogeneous training data across clients. We introduce \texttt{WAFFLE} (Weighted Averaging For Federated LEarning), a personalized collaborative machine learning algorithm that leverages stochastic control variates for faster convergence. \texttt{WAFFLE} uses the Euclidean distance between clients’ updates to weigh their individual contributions and thus minimize the personalized model loss on the specific agent of interest.

4: Through a series of experiments, we compare our new approach to two recent personalized federated learning methods---\texttt{Weight Erosion} and \texttt{APFL}---as well as two general FL methods---\texttt{Federated Averaging} and \texttt{SCAFFOLD}. Performance is evaluated using two categories of non-identical client data distributions---concept shift and label skew---on two image data sets (MNIST and CIFAR10). Our experiments demonstrate the comparative effectiveness of \texttt{WAFFLE}, as it achieves or improves accuracy with faster convergence.

5: \end{abstract}

6: