FedTSB: Tackling Sample Bias in Federated Services Selection

In recent years, federated learning (FL), a privacy-aware distributed learning paradigm, has been facilitating numerous cloud and mobile edge systems. Although FL proves effective in enhancing edge services, the complex circumstances of edge computing (e.g., data heterogeneity, limited communication capacity, and arbitrary service unavailability) can adversely affect FL applications, resulting in slow convergence and low-quality services. Existing FL methods merely address some of these issues, rendering them unsuitable for edge applications. To address these challenges, we propose a two-fold method, FedTSB, to alleviate both sustained bias and temporary bias, enabling rapid and stable training in complex edge computing scenarios. Experimental results on two datasets under three heterogeneous scenarios demonstrate that our proposed FedTSB outperforms four baseline methods in terms of both performance and efficiency metrics.