FedCSI: Profile-Driven Clustering and Cross-Scale Interaction for Heterogeneous Federated Graph Learning

Abstract

The proliferation of interconnected devices in the Internet of Things (IoT) has driven the adoption of federated graph learning (FGL) to extract relational intelligence without compromising edge privacy. However, deploying FGL in real-world IoT environments faces a critical bottleneck: the extreme statistical and structural heterogeneity across edge clients. Conventional global aggregation suffers from severe gradient conflicts, while existing clustered federated learning (CFL) methods struggle with unstable grouping due to their reliance on volatile model parameters. To overcome these limitations, we propose federated clustering via client self-inspection (FedCSI), a topology-aware framework adhering to a “Comprehend-Cluster-Optimize” protocol. Our approach begins with a self-supervised pretraining phase to establish a stable representation space, enabling clients to synthesize multidimensional profiles encompassing semantic, structural, and scale features without manual labels. The server leverages these profiles for precise, data-driven adaptive clustering. Within each cluster, a tailored cross-scale graph neural network (CrossScaleGNN) is deployed to explicitly capture fine-grained local interactions and macrolevel community topologies. Extensive experiments validate FedCSI’s effectiveness under strong heterogeneity ( $\beta =1.0$ ). It achieves up to a 2.4% absolute accuracy gain over standard FGL baselines. Notably, on the large-scale ogbn-arxiv topology under strong heterogeneity, FedCSI surpasses the standard centralized graph convolutional network (GCN) reference model (72.2% versus 70.0%). This result is attributed to the combined effect of reduced cross-client structural fragmentation, stronger semantic concentration under severe label skew, and CrossScaleGNN’s macrocommunity compensation. Furthermore, although FedCSI incurs additional communication cost due to pretraining and CrossScaleGNN parameter transmission, it provides an accuracy-oriented communication tradeoff and consistently outperforms representative FGL methods under strong heterogeneity. Overall, FedCSI establishes an efficient cloud-edge collaborative paradigm, delivering a favorable communication-to-accuracy tradeoff for resource-constrained IoT architectures. The code is publicly available at https://github.com/Xuan-Li1211/FedCSI-Pytorch.git