Clustering method based on data distribution similarity in personalized federation scene

• Main Authors: ZHOU YINING, XIANG TIAN'AO, BI YUANGUO, ZHI MINGJIAN
• Format: Patent
• Language: Chinese; English
• Published: 16.04.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; PHYSICS

The invention belongs to the technical field of personalization in a federated learning scene, and discloses a clustering method based on data distribution similarity in a personalized federated scene. Softmax output mean values of different user data under the same model approximately represent real data distribution of users, and the Wasserstein distance is calculated to obtain the similarity between the real data distribution. An improved two-stage hierarchical clustering method is adopted for clustering, and a cluster structure is determined. And finally, each user performs federal averaging in the own cluster, and updates the cluster head model of each cluster until the cluster head model converges. According to the method, the calculation overhead of the central server during execution of the clustering algorithm is reduced, the efficiency of determining the optimal cluster structure is improved, the robustness of the clustering algorithm is enhanced, and the problem of low clustering federal learning t