FedPredict: Combining Global and Local Parameters in the Prediction Step of Federated Learning

In traditional Federated Learning (FL), such as FedAvg, the main objective is to compute a generalized model applied to all clients. This approach is not effective in the non-IID scenario, where each client has a specific data distribution. As an alternative, personalized FL has proven to be an important research direction for dealing with clients' particularities. However, part of these solutions must be reexamined when a new client (i.e., a few times trained or never trained) is added to the FL process. To address these problems, we propose FedPredict, a simple but effective federated learning approach that combines global and local (i.e., personalized) model parameters of neural networks, considering their evolution and update levels. This combination is essential because our method is a plugin that operates in the prediction/inference step on the FL client side, which means that there is no modification in the learning process, and it can be coupled with other techniques. Compared to state-of-the-art solutions, FedPredict converges faster while achieving greater accuracy in various scenarios, including when new clients are added.