Boostfgl Achieves Fairness in Federated Graph Learning, Addressing 3 Key Disparities

Researchers are increasingly focused on ensuring fairness within federated learning, particularly when applied to complex graph-structured data. Zekai Chen, Kairui Yang, and Xunkai Li, from the Beijing Institute of Technology, alongside Henan Sun from The Hong Kong University of Science and Technology (GZ), and Zhihan Zhang and Jia Li, demonstrate that standard federated graph learning (FGL) methods can mask significant performance drops for specific, disadvantaged node groups despite achieving high average accuracy. This new work identifies three key sources of unfairness , label skew, topology confounding, and aggregation dilution , and introduces BoostFGL, a novel boosting framework designed to address these issues. By coordinating client-side node and topology boosting with server-side model boosting, BoostFGL demonstrably improves fairness, achieving an 8.43% increase in Overall-F1 across nine datasets, all while maintaining competitive overall performance , a crucial step towards equitable machine learning systems.

While existing FGL methods often achieve high overall accuracy, this research reveals that such performance can mask significant degradation experienced by disadvantaged node groups, particularly minority classes and heterophilous nodes. The team achieved this breakthrough by identifying three key, interconnected sources of unfairness: label skew towards majority patterns, topology confounding during message propagation, and aggregation dilution of updates from clients with challenging data. BoostFGL tackles these issues with a boosting-style approach, introducing coordinated mechanisms at the client and server levels to improve fairness without sacrificing overall performance.

Specifically, BoostFGL employs client-side node boosting to reshape local training signals, emphasizing under-served nodes and ensuring they receive adequate attention during learning. Simultaneously, client-side topology boosting reallocates propagation emphasis towards reliable structures and attenuates misleading neighbourhoods, improving the quality of information flow in sparse or heterophilous regions. Crucially, the framework also incorporates server-side model boosting, which performs difficulty- and reliability-aware aggregation, preserving informative updates from challenging clients while stabilising the global model, a significant advancement over uniform averaging techniques. Extensive experiments conducted across nine datasets demonstrate that BoostFGL delivers substantial fairness gains, improving Overall-F1 by 8.43% while maintaining competitive overall performance against strong FGL baselines.
The research establishes a new perspective on unfairness in FGL by decomposing biased training dynamics into these three core issues, label skew, topology confounding, and aggregation dilution, and linking them to specific disadvantaged node groups. This detailed analysis allows for targeted interventions at each stage of the client-server pipeline. BoostFGL’s modular design ensures compatibility with existing FGL methods and architectures, allowing it to be seamlessly integrated into current workflows and potentially stacked with other fairness-enhancing techniques. This work opens exciting possibilities for deploying fairer and more robust graph learning systems in sensitive applications such as recommendation systems, financial modelling, and biomedicine, where equitable outcomes are paramount.

Experiments show that BoostFGL not only improves fairness metrics but also maintains a 2.46% increase in overall Accuracy, demonstrating that fairness and performance are not mutually exclusive. The framework’s ability to handle large-scale graphs, where other competitive methods encounter memory limitations, further highlights its practical utility and scalability. By addressing the coupled issues of label imbalance, topological challenges, and server-side aggregation, BoostFGL represents a significant step towards building more equitable and effective federated graph learning systems for a wide range of real-world applications.

BoostFGL Addressing Fairness in Federated Graph Learning

Scientists developed BoostFGL, a boosting-style framework to address fairness issues in federated graph learning (FGL) . The research team demonstrated that existing FGL methods, while achieving high overall accuracy, can conceal significant performance degradation on disadvantaged node groups . This study pinpointed three coupled sources of disparity: label skew towards majority patterns, topology confounding during message propagation, and aggregation dilution of updates from challenging clients . To counteract these issues, the team engineered client-side node boosting, reshaping local training signals to prioritise systematically under-served nodes .

This technique effectively amplifies the influence of minority and hard-to-classify nodes during local model updates . Furthermore, scientists implemented client-side topology boosting, which reallocates propagation emphasis towards reliable, yet underused structures, while simultaneously attenuating misleading neighbourhoods . This innovative approach ensures more robust and accurate message passing, particularly in heterophilous or sparse regions of the graph . The study pioneered server-side model boosting, performing difficulty- and reliability-aware aggregation to preserve informative updates from clients with challenging data .

This method carefully weighs updates based on their difficulty and reliability, preventing the cancellation of valuable signals from hard clients by easier ones . Experiments employed nine datasets to rigorously evaluate BoostFGL’s performance, revealing a substantial fairness gain of 8.43% improvement in Overall-F1 . Importantly, BoostFGL maintained competitive overall performance compared to strong FGL baselines, demonstrating that fairness improvements did not come at the cost of overall accuracy . The team harnessed detailed process-level diagnostics to substantiate the amplification of disparities in federated settings, revealing skewed gradient allocation and unreliable message propagation . This work highlights the importance of a fairness-centred examination of FGL, considering not only average accuracy but also how that accuracy is achieved across different node groups and clients .

BoostFGL mitigates performance drops in federated learning

Scientists have developed BoostFGL, a new framework for fairness-aware federated learning (FGL) that addresses disparities in performance across different node groups. The research reveals that standard FGL methods, while achieving high overall accuracy, can conceal significant performance degradation on disadvantaged nodes, stemming from label skew, topology confounding, and aggregation dilution. Extensive experiments conducted on nine datasets demonstrate that BoostFGL improves Overall-F1 by 8.43%, while maintaining competitive overall performance compared to strong FGL baselines. Experiments revealed systematic label skew towards majority patterns, leading the team to implement client-side node boosting, which reshapes local training signals to emphasize under-served nodes.

Measurements confirm that this approach increases gradient-share parity, as formalized by Lemma 3.1, showing that increasing the boosting factor amplifies gradients from minority or hard nodes relative to the majority, pushing the gradient-share distribution towards a fairer baseline. The team measured Edge-wise Propagation Reliability (EPR), discovering that edges around minority and heterophilous regions exhibit a heavier negative tail, indicating harmful message passing. This structural propagation issue motivated the development of client-side topology boosting, which reallocates propagation emphasis towards reliable structures and attenuates misleading neighbourhoods, tests prove this suppresses error amplification with increasing hop distance, as illustrated in Figure 3. Data shows that standard averaging in server-side aggregation can cancel out fairness-improving updates, a phenomenon quantified by the Dilution Ratio (DR).

The researchers evaluated DR, finding it to be low under standard averaging, consistent with destructive interference among heterogeneous client updates. To address this, BoostFGL introduces server-side model boosting, performing difficulty- and reliability-aware aggregation to preserve informative updates from hard clients while stabilizing the global model. Theorem 3.3 formalizes that trust weights preserve minority-improving signals, matching the observed gains in DR, as depicted in Figure 0.15. Furthermore, theoretical analysis provides diagnostic-aligned guarantees, demonstrating that the three boosting modules in BoostFGL monotonically improve process-level signals: Gradient-Share Distribution (GSD), EPR, and DR.

Specifically, node-side boosting increases GSD, topology-side boosting suppresses harmful messages (negative EPR), and model-side boosting mitigates aggregation dilution. The team established that in a high-confidence stationary regime, the boosting factors vanish, and the procedure converges to standard FedAvg, ensuring asymptotic consistency. These findings deliver a substantial advancement in fairness-aware FGL, offering a robust and effective solution for mitigating performance disparities in decentralized learning environments.

BoostFGL mitigates fairness issues in federated learning

Scientists have demonstrated that federated learning (FGL) methods, while achieving high overall accuracy in graph neural networks (GNNs), can mask significant performance degradation affecting disadvantaged node groups. Researchers identified three interconnected causes for these fairness issues: label skew towards common patterns, topological confounding during message propagation, and dilution of updates from challenging clients. To remedy this, they developed BoostFGL, a boosting-style framework designed for fairness-aware FGL. BoostFGL incorporates three coordinated mechanisms: client-side node boosting to prioritise under-served nodes, client-side topology boosting to focus on reliable network structures, and server-side model boosting to intelligently aggregate updates from difficult clients while maintaining global model stability.

Experiments conducted across nine datasets revealed that BoostFGL substantially improves fairness, increasing Overall-F1 by 8.43%, all while maintaining competitive overall performance compared to existing FGL approaches. The framework also exhibited robust behaviour with varying hyperparameters and favourable efficiency. The authors acknowledge that their work focuses on diagnosing and mitigating unfairness at specific stages of the FGL pipeline, rather than solely relying on a global fairness objective. Future research could explore the application of BoostFGL to more complex graph structures and investigate its compatibility with diverse privacy-preserving techniques, such as differential privacy, demonstrated through simulations showing robustness to DP-style noise injection. These findings suggest a practical design principle for fairness-aware FGL: addressing unfairness with stage-specific signals and mitigating it at the source, rather than through a global objective, is a promising avenue for development.