Abstract
Federated node classification is not easy when clients hold induced subgraphs with different class proportions and different local topology. Existing graph-specific approaches change aggregation, distillation or cross-client knowledge transfer, which can increase personalization but can also be associated with a different global optimization path. We introduce a simple two-stage method that first trains and validation-selects a shared graph convolutional network using FedAdagrad and freezes its encoder, and then each client will only update the prediction head of the graph convolutional network based on local learning data and local validation early on. Personalized heads are never uploaded, so the core method has no communication beyond the global model. A global parity analysis shows that the accuracy-trajectory gap is identical for the FedAdagrad and Stage-I models. Our method achieves the highest mean accuracy on Cora (83.14%), CiteSeer (75.69%), and PubMed (90.66%) compared to the best generic baseline for each dataset. The paired gains are +0.20 pp on Cora, +0.87 pp on CiteSeer, and +2.20 pp on PubMed; the latter two can be obtained after Holm correction for the same dataset, while Cora cannot. For the same FedAdagrad checkpoint, the PubMed gain is +3.67 pp at no further communication. For a six-level Dirichlet sweep, we observe a strong increase in the average label bias on CiteSeer and PubMed, to 1.83 pp and 6.42 pp, respectively. Our method is not destructive with Louvain and METIS partitioning, is very robust to 5–20 clients and 50–100% participation, and improves ogbn-arxiv accuracy from 41.51% to 52.27% in a five-seed federated scalability stress test using the same language. Ablations show that training-only prototypes and validation-gated ensembling do not consistently improve upon head-only adaptation. A matched personalization suite further shows that FedPer/FedRep can be more accurate on Cora and CiteSeer, while FedColMem-H remains close to FedPer and competitive with FedRep on PubMed, clarifying the method’s accuracy–intervention trade-off. A separate backbone-transfer experiment finds positive mean gains in all 18 GCN/GraphSAGE/GAT dataset–heterogeneity conditions, with the largest effects on PubMed. The ogbn-arxiv number is from a custom federated repartitioning stress test and is not an official OGB leaderboard result. These results suggest that when a useful shared graph representation can be learned globally, post-hoc local decision-boundary adaptation is an excellent communication-neutral complement to federated optimization.