<p>Federated Learning (FL) is a learning paradigm in which clients are trained in a decentralized manner on their local datasets. Each client aims to optimize its individual objective while leveraging the advancements made by other clients in the network, without explicitly sharing its local data to ensure user privacy. Heterogeneous data distribution refers to situations where a client or cluster of clients have data whose data distribution significantly differs from that of others in the network; in such cases, typical federated learning algorithms may struggle to perform effectively. In our approach, we follow the personalized federated learning (PFL) framework and develop personal local models rather than aiming for global or cluster models. We propose a similarity-based algorithm, FedKNN, which utilizes similarity between clients as part of the aggregation and restricts aggregation to the most similar neighborhood of the client. We examine our algorithm’s performance in a variety of heterogeneous data scenarios simulated via dirichlet sampling and pathological sampling. We also consider the scenario of data drift, where clients’ data evolves over time, which harms clients’ performance. We evaluate our proposed approach in severe data drift scenarios and compare it with existing baseline FL algorithms. We represent the client networks as graph and show that through similar neighborhood aggregation clients can benefit from the collective knowledge of the network and have flexibility in adapting to data drift. We highlight and provide insights into the behavior and performance gains of our algorithm, particularly, we highlight faster convergence and improvement of average test accuracy under heterogeneous scenarios and reduced fluctuations in data drift scenario compared to baseline algorithms.</p>