AIR: Adaptive Incremental Embedding Updating for Dynamic Knowledge Graphs

Recently, knowledge graphs (KGs) have been successfully applied on various downstream tasks, such as online recommendation and question answering. To better utilize the knowledge in KGs, recent researchers propose to learn low-dimension embeddings for entities and relations in KGs. Unlike static KGs, dynamic KGs evolve with knowledge events. Therefore, the embeddings of entities and relations should be updated to capture the information of latest knowledge events. However, when new events arrive, existing approaches update the whole knowledge graph or only update limited neighbors of entities directly involved in events. These approaches cannot maintain high-quality KG embedding in an efficient way. In this paper, we propose a framework AIR to adaptively update the embeddings for dynamic KGs. Specifically, AIR first measures the importance score of each triple. Then, AIR selects a set of triples that are most affected by knowledge events for updating. Besides, we apply the embedding propagation method to update the embedding to avoid retraining. Therefore, our proposed AIR can efficiently maintain high-quality KG embeddings. Extensive experiments on four real-world dynamic KGs demonstrate the effectiveness and efficiency of AIR against state-of-the-art baselines.