Research map / Federated learning and privacy / Personalized and heterogeneous FL
Towards Fair Graph Federated Learning via Incentive Mechanisms
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an incentive mechanism for graph federated learning that allocates sparsified model gradients by gradient-alignment plus motif-based graph-diversity agent values, and payoff with punishment and delay compensation, plus motif prototypes to improve accuracy.
Problem
Graph federated learning agents are self-interested; existing FL incentive mechanisms assume all agents contribute positively and promptly, ignoring harmful agents and delayed contributors in graph settings.
Approach
Combines model-gradient allocation (sparsified global gradients scaled by agent value) with payoff allocation (punishment for negative value, post-hoc compensation for delayed contributors); agent value uses gradient-based Shapley approximation (cosine similarity) and motif-based graph diversity; motif prototypes are shared between server and agents to align local training and aggregation.
Key results
- On PROTEINS, the method reaches model gradient fairness 0.787 and global accuracy 0.753, the best among compared methods.
- On IMDB-BINARY, model gradient fairness reaches 0.907 versus 0.874 for the best baseline (GCFL).
- In payoff experiments with perturbed data, the method gives high-quality agents the highest payoff and assigns negative payoff to low-quality agents.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add gradient alignment to agent valuation (ablation) | FairGraphFL · PROTEINS | model gradient fairness: 0.736±0.145 -> 0.787±0.052 | Helped |
| Add graph diversity to agent valuation (ablation) | FairGraphFL · PROTEINS | model gradient fairness: 0.723±0.082 -> 0.787±0.052 | Helped |
| Add model gradient allocation (ablation) | FairGraphFL · PROTEINS | model gradient fairness: 0.710±0.146 -> 0.787±0.052 | Helped |
| Use value-based rather than FedAvg aggregation (ablation) | FairGraphFL · PROTEINS | personalized accuracy: 0.741±0.042 -> 0.751±0.017 | Helped |
| Add motif prototype components (ablation) | FairGraphFL · PROTEINS | personalized accuracy: 0.737±0.025 -> 0.751±0.017 | Helped |
Surprising findings
- Some agents in graph FL make negative contributions even though all participants are honest, unlike image FL where all agents contribute positively.
- Removing the model-gradient allocation component improves personalized accuracy (0.757 vs 0.751) while hurting gradient fairness, a trade-off the authors highlight.
Limitations
- Only honest agents are considered; potential cheating behaviors are not examined.
- Evaluation limited to graph classification on small molecular and social graph datasets with 10 agents.
- Accuracy-fairness trade-off controlled by hyperparameter beta; personalized accuracy is sometimes lower than the best baselines (e.g., GCFL on DD and IMDB-BINARY).
Experimental setup
Datasets: PROTEINS DD IMDB-BINARY TWITTER-Real-Graph-Partial
Benchmarks: PROTEINS DD IMDB-BINARY TWITTER-Real-Graph-Partial
Metrics: model gradient fairness (Pearson correlation) global accuracy personalized accuracy payoff fairness
Baselines: Self-train FedAvg FedProx FedSage GCFL GCFL+ FedStar DW EU CFFL equal incentive union incentive
Builds on: FedAvg Shapley value GCFL FedStar
Closest papers by meaning
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.88)
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.86)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.86)
- Subgraph Federated Learning via Spectral Methods (NeurIPS 2025 · similarity 0.86)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 2024 · similarity 0.86)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.