Federated Learning FedLTailor: A Dynamic Weight Adjustment and Personalized Fusion Approach

In this study, we primarily address the issue of uneven quality of client embeddings in existing federated learning frameworks for knowledge graph completion. Although existing frameworks provide preliminary solutions to the heterogeneity of data in knowledge graphs, there are still deficiencies in their aggregation strategies. To address this issue, we introduce a new federated learning framework called FedLTailor that focuses on optimizing the aggregation process. FedLTailor employs a dynamic weight adjustment strategy to enhance the weight proportion of clients with higher embedding quality during the aggregation process, thereby optimizing the performance of the global model. Moreover, FedLTailor adopted a unique personalized fusion strategy to mitigate potential discrepancies between the global model and local clients. Experimental results on four federated knowledge graph datasets demonstrate FedLTailor's significant advantages in addressing aggregation issues in federated knowledge graph completion tasks, as well as its broad adaptability across various knowledge graph embedding techniques. Additionally, the design and experimental validation of the FedLTailor framework offers new insights into the field of federated learning, particularly in handling distributed knowledge graph data, showcasing its potential applicability and effectiveness.