Federated Learning with Partially Class-Disjoint Data
Ziqing Fan, Jiangchao Yao · Artificial intelligence · 2024
Essentially, the federation of multiple clients is to acquire more complete information to promote learning a powerful model that recognizes general patterns of wider classes. However, a natural case usually presents in the real-world situation under a large label space, where only a subset of classes of samples are provided by individual clients. We term this practical setting as “federated learning with partially class-disjoint data” (PCDD), to differentiate from the ordinary study without emphasizing this extreme scenario. Generally, this setting is very challenging as different clients own partial classes and the resulting model on the client side cannot well support full-class topology for aggregation. In this chapter, we first show the significant difference of federation in PCDD compared to that in traditional non-IID studies, and then dissect recent advances concerning this challenge, following with discussion about the future trends. We illustrate the challenges using experimental results on benchmark datasets under PCDD conditions, finding that effective solutions to PCDD problems should provide optimal structures for global objective and allow flexible structures for personalized tasks. We make systematical analysis to underscore the significance of PCDD in heterogeneous federated learning and to inspire more explorations in the future.