Federal Knowledge Graph Embedding Based on Incentive Mechanism

Lijun Chen, Yudong Zhang, Xia Xie, Mianxiong Dong, Kaoru Ota

Abstract: Currently, the task of knowledge graph inference based on federated learning remains impractical, primarily because data controllers are reluctant to share their data. Traditional knowledge graph embedding methods based on consortium learning can fully utilize multi-source knowledge graph data while ensuring that local data is not leaked. However, these traditional approaches depend on the voluntary participation of data controllers in federated learning, which poses a significant challenge in practice. Encouraging the active participation of data controllers is a key issue. This paper proposes a federated learning framework for entity embedding in multi-source knowledge graphs and introduces an incentive mechanism designed to increase participation through an incentive algorithm. Additionally, a new dataset is constructed based on mainstream public datasets. Experimental results demonstrate that the proposed method achieves superior performance compared to baseline methods.