GraphFedMIG: Tackling Class Imbalance in Federated Graph Learning via Mutual Information-Guided Generation

Xinrui Li, Qilin Fan, Tianfu Wang, Kaiwen Wei, Ke Yu, Xu Zhang

TL;DR

GraphFedMIG solves the critical issue of class imbalance in Federated Graph Learning (FGL) by reframing it as a generative data augmentation task. By combining a hierarchical GAN architecture with a mutual information-guided aggregation mechanism, it synthesizes high-quality minority-class features while maintaining training stability and data privacy.

What Problem Does This Paper Address?

Federated Graph Learning (FGL) allows multiple institutions to collaboratively train Graph Neural Networks (GNNs) without centralizing their private, sensitive graph data. However, real-world data is rarely distributed evenly. A major hurdle in FGL is class imbalance—for example, in financial fraud detection, legitimate transactions vastly outnumber fraudulent ones.

Standard FGL models trained on such data become heavily biased toward the majority class. In graph structures, this problem is amplified because minority nodes are often surrounded by majority-class neighbors, drowning out rare but critical signals during GNN message passing. Existing solutions typically rely on model-centric fixes like loss re-weighting, which only patch over the problem rather than fixing the root cause: data scarcity. Furthermore, simply applying Generative Adversarial Networks (GANs) in a naive federated setup leads to severe training instability and mode collapse due to the destructive nature of standard weight averaging (FedAvg) on adversarial components.

Preliminary experiments showing the failure of naive GAN integration with FedAvg

Methods

The authors introduce GraphFedMIG (Mutual Information-Guided Generation), a novel framework that treats class imbalance as a federated generative data augmentation task. Its architecture consists of three main pillars:

  1. Hierarchical GAN Architecture: To balance computational efficiency and personalization, GraphFedMIG groups clients into clusters using Agglomerative Hierarchical Clustering based on their initial local data distributions. Each client trains a local generator to synthesize node embeddings, while clients within the same cluster share a dedicated discriminator to provide tailored supervision.
  2. Privacy-Preserving Feature Prototypes: Instead of sharing raw node features to train the cluster discriminators, clients compute and share low-dimensional class centroids (prototypes). These prototypes act as a secure proxy for the true data distribution and dynamically co-evolve as the GNNs improve.
  3. Mutual Information-Guided Model Aggregation (MIGMA): To counteract aggregation bias, the server calculates each client's unique informational value using Jensen-Shannon (JS) divergence. Instead of standard averaging, clients perform a local parameter correction ($\tilde{\theta}_m = W_m \theta_m$) based on these weights, ensuring that clients holding rare, high-value minority data exert a stronger influence on the generative process.
Figure 2: (a) An overview of the proposed GraphFedMIG, a hierarchical GAN framework for FGL. It features client-side generators supervised by cluster-specific discriminators. The framework includes a mutual information-guided method for correcting generator parameters and a cross-cluster aggregation strategy on the server for updating discriminator parameters. (b) An illustration of the detailed workflow and the specific loss functions that drive the model.

Datasets

The framework was evaluated on four real-world, class-imbalanced graph datasets partitioned to simulate a distributed federated setting:

  • Elliptic: A Bitcoin transaction network for anomaly detection (26.6% minority/illicit transactions).
  • Twitch: A social network of streamers used to detect policy violations (38.9% minority).
  • FacebookPagePage (Facebook): A network of official pages categorized for user recommendation (15.1% minority).
  • Actor: A co-actor network derived from film collaborations for multi-class actor classification (11.0% minority).

Key Insights and Ideas

  • Data-Level Augmentation Wins: Addressing minority data scarcity at the source by generating high-fidelity synthetic features is fundamentally more robust than trying to compensate via model parameter tweaks.
  • Mutual Information as a Regularizer: Weighting client updates based on mutual information acts as an anchor that stabilizes adversarial training. As shown in the paper's convergence analysis, it prevents the volatility typically seen in federated GANs.
  • Cooperative Competition: The framework's composite loss function incentivizes local generators to not only fool their cluster discriminator but also align their output distributions with their cluster peers, leading to a more thorough and structured exploration of the feature space.

Key Results

GraphFedMIG consistently outperformed state-of-the-art baselines (including FedAvg and FedSpray), showing massive gains in minority-class metrics:

  • Minority Recall Boost: On the Facebook dataset, GraphFedMIG achieved a 64.8% minority recall, representing a 13.09% improvement over the next best baseline.
  • Faster Convergence: GraphFedMIG reached a stable performance plateau around the 50th communication round, whereas variants without mutual information guidance required up to 90 rounds and suffered from high volatility.
Convergence curves demonstrating the stabilizing effect of GraphFedMIG components
  • Clearer Class Separation: 2D scatter plot visualizations confirmed that the synthetic data generated by GraphFedMIG formed much more distinct, well-separated clusters compared to standalone GANs.
2D visualization of data distributions showing superior class separation by GraphFedMIG

Limitations and Future Work

  • Communication Overhead: Exchanging generator parameters, discriminators, and feature prototypes increases the communication payload per round (0.089 MB for GraphFedMIG vs. 0.049 MB for standard FedAvg). While the authors kept these components lightweight to ensure they operate within the same order of magnitude, scaling to massive GNN architectures could challenge bandwidth-constrained clients.
  • Clustering Dependency: The client clustering happens as a one-time pre-processing step based on initial local GNN training. If a client's local data is too scarce to train a reasonable initial model, it might be assigned to an suboptimal cluster, potentially degrading downstream generation quality.

Summary updated

Abstract

Federated graph learning (FGL) enables multiple clients to collaboratively train powerful graph neural networks without sharing their private, decentralized graph data. Inherited from generic federated learning, FGL is critically challenged by statistical heterogeneity, where non-IID data distributions across clients can severely impair model performance. A particularly destructive form of this is class imbalance, which causes the global model to become biased towards majority classes and fail at identifying rare but critical events. This issue is exacerbated in FGL, as nodes from a minority class are often surrounded by biased neighborhood information, hindering the learning of expressive embeddings. To grapple with this challenge, we propose GraphFedMIG, a novel FGL framework that reframes the problem as a federated generative data augmentation task. GraphFedMIG employs a hierarchical generative adversarial network where each client trains a local generator to synthesize high-fidelity feature representations. To provide tailored supervision, clients are grouped into clusters, each sharing a dedicated discriminator. Crucially, the framework designs a mutual information-guided mechanism to steer the evolution of these client generators. By calculating each client's unique informational value, this mechanism corrects the local generator parameters, ensuring that subsequent rounds of mutual information-guided generation are focused on producing high-value, minority-class features. We conduct extensive experiments on four real-world datasets, and the results demonstrate the superiority of the proposed GraphFedMIG compared with other baselines.
BibTeX
@misc{li2025graphfedmig,
  author = {Xinrui Li and Qilin Fan and Tianfu Wang and Kaiwen Wei and Ke Yu and Xu Zhang},
  title = {{GraphFedMIG: Tackling Class Imbalance in Federated Graph Learning via Mutual Information-Guided Generation}},
  year = {2025},
  eprint = {2508.10471},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.10471},
}
Figure