Federated Continual Graph Learning

TL;DR

POWER is introduced, a framework that preserves experience nodes with maximum local-global coverage locally to mitigate LGF, and leverages pseudo-prototype reconstruction with trajectory-aware knowledge transfer to resolve GEC.

Abstract

Managing evolving graph data presents substantial challenges in storage and privacy, and training graph neural networks (GNNs) on such data often leads to catastrophic forgetting, impairing performance on earlier tasks. Despite existing continual graph learning (CGL) methods mitigating this to some extent, they rely on centralized architectures and ignore the potential of distributed graph databases to leverage collective intelligence. To this end, we propose Federated Continual Graph Learning (FCGL) to adapt GNNs across multiple evolving graphs under storage and privacy constraints. Our empirical study highlights two core challenges: local graph forgetting (LGF), where clients lose prior knowledge when adapting to new tasks, and global expertise conflict (GEC), where the global GNN exhibits sub-optimal performance in both adapting to new tasks and retaining old ones, arising from inconsistent client expertise during server-side parameter aggregation. To address these, we introduce POWER, a framework that preserves experience nodes with maximum local-global coverage locally to mitigate LGF, and leverages pseudo-prototype reconstruction with trajectory-aware knowledge transfer to resolve GEC. Experiments on various graph datasets demonstrate POWER's superiority over federated adaptations of CGL baselines and vision-centric federated continual learning approaches.

Figures (7)

• Figure 1: Illustration of our FCGL paradigm, which achieves collaborative CGL through multi-round GNN parameter communications between clients and server. Each client's evolving graph is represented by two tasks, with node colors indicating different categories.
• Figure 2: Experimental results of our empirical study. (a) Node label distribution of each client's class-incremental tasks, each exhibiting divergent evolution trajectories. (b) Comparative analysis of three CGL methods in isolated versus federated settings, presenting AM and FM metrics in the upper and lower parts, respectively. (c) Performance variation of Client 1's local GNN (Single GNN) during training on Client 1 Task 3, with markers denoting receiving global parameters from the server. (d) Performance comparison between clients and the server (Multiple GNNs) during training on Client 1 Task 3, indicating that the global GNN can be improved by local GNNs with expertise.
• Figure 3: Overview of our proposed POWER framework. We follow a class-incremental setting, where each client collects an evolving graph into which unseen classes are continually introduced. Node colors indicate different labels. (1) To tackle LGF, POWER stores experience nodes with maximum local-global coverage in the last round of the old task and replays them in each round of the new task; (2) To tackle GEC, POWER reconstructs pseudo prototypes via client-server collaboration in the first round of each task. Based on this, POWER constructs a global buffer graph and applies trajectory-aware knowledge transfer to recover the global GNN’s lost knowledge in each round of each task.
• Figure 4: Sensitivity analysis for the replay buffer size $b$ across eight benchmark graph datasets.
• Figure 5: Sensitivity analysis for various combinations of the trade-off parameters $\beta$ and the decay coefficient $\phi$ on the Cora dataset.