HFKG-RFE: An algorithm for heterogeneous federated knowledge graph.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated learning enables global model training without local data sharing, crucial for sensitive sectors like healthcare and finance.
• Knowledge graphs represent information akin to human cognition, but their training demands substantial data.
• Combining federated learning with knowledge graphs presents challenges due to statistical heterogeneity in federated data, impacting model performance.

Purpose of the Study:

• To address the challenge of statistical data heterogeneity in federated knowledge graph training.
• To propose novel solutions for improving the performance of federated knowledge graph embedding models.

Main Methods:

• Introduced the Heterogeneous Federated Knowledge Graph (HFKG) algorithm, employing comparative learning to mitigate model drift.
• Developed a new server aggregation algorithm and a knowledge graph embedding model named RFE.
• Conducted extensive experiments using DDB14, WN18RR, and NELL datasets with varied partitioning strategies to simulate heterogeneity.

Main Results:

• Demonstrated stable performance improvements across experiments.
• Validated the effectiveness of the HFKG-RFE algorithm for federated knowledge graph embedding aggregation.
• Confirmed the efficacy of the RFE knowledge graph embedding model.

Conclusions:

• The proposed HFKG-RFE algorithm and RFE model effectively overcome statistical data heterogeneity in federated knowledge graph learning.
• The combined approach significantly enhances the overall performance of federated knowledge graph embedding aggregation.