Scholarlink

<p>Federated learning provides privacy protection for distributed high-performance computing to meet the demands of large-scale data environments. The existing work neglects the connections between participants during collaborative training, resulting in the compromised performance of federated learning on heterogeneous graph data. To tackle this issue, we introduce a novel federated learning algorithm that constructs a participant graph and broadcasts parameters across it without introducing privacy risks. This algorithm restores implicit connections between participants by broadcasting parameters. The participant graph is initially constructed from the extended nodes generated by each participant, and an alliance is formed by permuting the ordered participant nodes on the graph. Subsequently, parameters are broadcast to other participants within an alliance, and participants’ contribution weights are determined based on the marginal revenue concept in Shapley’s theory. The global model is finally obtained through weighted aggregation. Extensive experiments are conducted to validate the effectiveness and robustness of our method for enhancing the performance of federated learning on heterogeneous graph data. Our method achieves 1.54%, 1.77%, and 1.36% improvements in accuracy over FedGCN using GCN on Cora, Citeseer, and PubMed, respectively. We also validate this method on two real-world financial risk datasets to illustrate its potential in engineering applications. The testing accuracy of our method with GraphSAGE is 63.91% and 63.45% on Elliptic and EthereumHeist, which is 2.87% and 3.62% higher than FedEgo.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Parameter broadcasting on participant graph for federated heterogeneous graph learning

  • 期刊论文

  • 发表日期：

    2026年5月16日

查看全文
  • Juncheng Pu,
  • Xiaodong Fu,
  • Li Liu,
  • Lianyin Jia,
  • Jie Li

摘要

Federated learning provides privacy protection for distributed high-performance computing to meet the demands of large-scale data environments. The existing work neglects the connections between participants during collaborative training, resulting in the compromised performance of federated learning on heterogeneous graph data. To tackle this issue, we introduce a novel federated learning algorithm that constructs a participant graph and broadcasts parameters across it without introducing privacy risks. This algorithm restores implicit connections between participants by broadcasting parameters. The participant graph is initially constructed from the extended nodes generated by each participant, and an alliance is formed by permuting the ordered participant nodes on the graph. Subsequently, parameters are broadcast to other participants within an alliance, and participants’ contribution weights are determined based on the marginal revenue concept in Shapley’s theory. The global model is finally obtained through weighted aggregation. Extensive experiments are conducted to validate the effectiveness and robustness of our method for enhancing the performance of federated learning on heterogeneous graph data. Our method achieves 1.54%, 1.77%, and 1.36% improvements in accuracy over FedGCN using GCN on Cora, Citeseer, and PubMed, respectively. We also validate this method on two real-world financial risk datasets to illustrate its potential in engineering applications. The testing accuracy of our method with GraphSAGE is 63.91% and 63.45% on Elliptic and EthereumHeist, which is 2.87% and 3.62% higher than FedEgo.