Temporal Knowledge Graph Embedding Via Sparse Transfer Matrix
Xin Wang,Shengfei Lyu,Xiangyu Wang,Xingyu Wu,Huanhuan Chen
DOI: https://doi.org/10.1016/j.ins.2022.12.019
IF: 8.1
2023-01-01
Information Sciences
Abstract:Knowledge Graph Completion (KGC) is a fundamental problem for temporal knowledge graphs (TKGs), and TKGs embedding methods are one of the essential methods for KGC. However, existing TKG embedding methods encounter a scalability dilemma, i.e., the inconsistency in parameter scalability among different datasets, and the less use of global information, e.g., statistics and dependencies of facts. To mitigate these two issues, we propose a novel and effective TKG embedding method, named T emporal Knowledge Gr a ph Embedding via S parse T ransf e r Mat r ix (TASTER), which provides a framework to utilize both global and local information. Regarding a TKG as a static knowledge graph when ignoring the time dimension, TASTER first learns global embeddings based on this static knowledge graph to capture global information. To capture the local information in a specific timestamp, TASTER evolves local embeddings from global embeddings based on the corresponding subgraph. Besides, TASTER learns evolving entity embeddings through sparse transformation matrices, which could better adapt to TKGs with a varied number of subgraphs. We conduct experiments on three real-world datasets, and TASTER outperforms most existing models on the link prediction task of TKGs, which validates the its effectiveness.