Federated Graph Neural Networks under Non-IID Client Partitions
Nusrat Jahan
Jahangirnagar University
Abstract
Federated learning over graph-structured data is complicated by the fact that a client's local subgraph is rarely a uniform sample of the global graph. We characterise the degradation of FedAvg-style aggregation under degree-skewed and community-skewed partitions, and introduce a neighbourhood-reweighting scheme that recovers 94% of centralised accuracy on three citation benchmarks while exchanging no node features. We further show that the reweighting term admits a closed-form estimate from degree histograms alone, so clients need not disclose their subgraph topology. Experiments on Cora, CiteSeer and PubMed under six partition regimes support the analysis.
Keywords:
federated learninggraph neural networksprivacynon-IID datadistributed systems
References
- Kipf, T. N., & Welling, M. (2017). Semi-supervised classification with graph convolutional networks. ICLR.
- McMahan, B., et al. (2017). Communication-efficient learning of deep networks from decentralized data. AISTATS.
- Hamilton, W., Ying, Z., & Leskovec, J. (2017). Inductive representation learning on large graphs. NeurIPS.
- Zhao, Y., et al. (2018). Federated learning with non-IID data. arXiv:1806.00582.