APFedEmb: An Adaptive and Personalized Federated Knowledge Graph Embedding Framework for Link Prediction

Tengfan Weng, Xiaoyu Kang, Zhixin Shi

Abstract: Federated Knowledge Graphs (FKGs) seek to utilize dispersed knowledge while maintaining data privacy. The differences in data among clients and the need for custom models create challenges for existing FKG embedding methods in predicting links. This research presents APFedEmb (Adaptive and Personalized Federated Embedding), an innovative framework for link prediction in FKGs. APFedEmb tackles data heterogeneity by employing adaptive weight aggregation and optimizing tailored client embeddings. It utilizes a loss-aware weight aggregation approach that dynamically modifies aggregate weights according to client model performance, thereby prioritizing models trained on superior data quality. Additionally, APFedEmb allows each client to optimize their own embeddings, enabling changes to hyperparameters to better match their local data characteristics. In-depth research on standard FKGs datasets like FB15k-237 and WN18RR shows that APFedEmb performs much better than existing FKG embedding methods when it comes to predicting links. APFedEmb enhances the quality of FKGs embeddings and link prediction results in federated settings by effectively handling different types of data with smart combining and customized improvements.

External IDs:dblp:conf/icic/WengKS25