FedE: Embedding Knowledge Graphs in Federated Setting

Knowledge graphs (KGs) become widespread and many organizations construct as well as maintain their own knowledge graphs. Same as the data isolation which has been a long-standing problem, knowledge graph isolation is common in real knowledge graph applications. Since the incompleteness of knowledge graphs obtained by different owners, they need to take advantage of other knowledge graphs to complete their own knowledge graphs, without exposing knowledge graphs explicitly since the consideration of data privacy, commercial interests and so on. Knowledge graph embedding (KGE) methods represent components of a knowledge graph as vectors in continuous vector spaces (i.e., embeddings) and proved to be effective in conducting knowledge graph completion. However, current knowledge graph embedding methods focus on the scenario with only a single knowledge graph. To solve this problem, we introduce the federated setting for knowledge graphs and apply it in knowledge graph embedding. We propose a Federated Knowledge Graph Embedding framework, FedE, focusing on learning knowledge graph embeddings by aggregating locally-computed updates. In this framework, there is a client for each knowledge graph and a server for coordinating embedding aggregation. Specifically, entity embeddings are locally learned in clients and the server is responsible for aggregating entity embeddings from clients. Furthermore, a model fusion procedure blends the capability of learned embeddings based only on one client without using the federated setting and embeddings based on all the clients in the federated setting. Finally, we conduct extensive experiments on datasets derived from KGE benchmark datasets, and results show the effectiveness of our proposed FedE.