Abstract
Federated Graph Learning (FGL) enables collaborative training of Graph Neural Networks (GNNs) across decentralized clients while preserving data privacy. However, real-world graph data often suffers from extreme class imbalance, where minority classes represent critical but rare events. Standard FGL algorithms often fail to capture these signals, biasing the global model toward majority classes. We present GraphFedMIG, a framework that addresses this by reframing imbalance as a generative data augmentation task. GraphFedMIG utilizes a hierarchical GAN architecture and a Mutual Information-Guided (MIG) mechanism to synthesize high-quality minority-class embeddings. Experimental results on real-world datasets demonstrate that GraphFedMIG significantly improves minority class detection and convergence speed compared to state-of-the-art baselines.
Introduction
Federated Graph Learning (FGL) is an emerging paradigm for training GNNs on distributed datasets, such as financial transaction networks or medical records, without centralizing sensitive information. A major challenge in this domain is class imbalance. In applications like fraud detection or rare disease diagnosis, the "interesting" nodes are often significantly outnumbered by the majority class. Standard federated aggregation methods like FedAvg tend to marginalize these rare signals. Furthermore, the structural dependencies in graphs exacerbate this, as minority nodes are frequently embedded within majority-class neighborhoods, making it difficult for models to learn meaningful representations.
Methods
GraphFedMIG introduces a data-centric approach to FGL imbalance by generating synthetic minority-class data locally. Instead of merely re-weighting loss functions, the framework creates new node embeddings to fill data gaps.
The framework consists of several key innovations:
- Hierarchical GAN Architecture: To manage computational costs and data heterogeneity, clients are clustered based on data distributions. Each cluster shares a discriminator on the server, while clients maintain local generators to synthesize node embeddings.
- Privacy-Preserving Prototypes: Instead of sharing raw data, clients communicate class-wise centroids (prototypes) to guide the GAN. These act as a privacy-safe proxy for the global data distribution.
- Mutual Information-Guided (MIG) Mechanism: This mechanism uses Jensen-Shannon divergence to calculate the "informational value" of each client. This value is used to weight model aggregation and "correct" local generator parameters, ensuring the model prioritizes high-quality, unique information from minority classes.
- Composite Loss Function: Generators are optimized using a combination of classification loss, adversarial loss (to fool the discriminator), and a Mutual Information Fidelity Loss (using InfoNCE) to ensure synthetic data remains grounded in real-world characteristics.
Key Results
GraphFedMIG was evaluated on four real-world, class-imbalanced graph datasets: Elliptic (Bitcoin fraud), Twitch (user behavior), Facebook (recommendations), and Actor (collaboration). The framework consistently outperformed state-of-the-art baselines like FedAvg and FedSpray.
Key findings include:
- Minority Accuracy Boost: On the Facebook dataset, GraphFedMIG improved minority class accuracy by 13.95% over the best baseline.
- Superior Recall: The framework achieved higher overall and minority recall in almost all scenarios, proving its effectiveness at identifying rare but critical events.
- Training Stability and Convergence: The MIG mechanism acts as a regularizer, preventing GAN mode collapse. In the Elliptic experiments, GraphFedMIG reached stable performance in approximately 50 rounds, significantly faster than simpler GAN-augmented versions which remained volatile after 90+ rounds.
- Regularization via Privacy: Experiments with Differential Privacy (DP) revealed that moderate noise ($\epsilon = 0.1$) actually improved performance, suggesting that DP noise can act as a form of regularization that prevents the generator from overfitting to local data.
Conclusion
GraphFedMIG effectively addresses the class imbalance problem in FGL by shifting the focus from model-centric loss weighting to data-centric generative augmentation. By leveraging mutual information to guide the generation process, it provides a robust and privacy-preserving solution for detecting rare events in decentralized graph networks. While the framework introduces some communication overhead and implementation complexity, its gains in minority class detection and training stability offer a significant advancement for real-world federated applications.
Summary updated