Personalized federation algorithm based on model similarity and data importance

As an emerging distributed machine learning approach, federated learning has gradually become an effective method for addressing the issue of data silos. However, traditional federated learning algorithms often fail to achieve ideal personalization effects when dealing with data distribution of different clients, thus affecting the overall performance of the model. To solve this problem, this paper proposes a personalized federated learning algorithm FedSIA based on model similarity and data importance. FedSIA calculates the cosine similarity between client models, constructs a similarity matrix, and evaluates the importance of clients by combining data volume and data quality, thereby redistributing weights in the model aggregation stage. This paper also introduces L2 regularization in the client model update process to ensure the balance between personalized models and historical models. Experimental results show that FedSIA performs well on MNIST and CIFAR-10 datasets, especially in data heterogeneous scenarios, where FedSIA achieves higher accuracy, outperforming traditional federated learning algorithms and other classic personalization methods.