T-GAE: A Timespan-aware Graph Attention-based Embedding Model for Temporal Knowledge Graph Completion

• Published in: Information sciences Vol. 642; p. 119225
• Main Authors: Hou, Xiangning, Ma, Ruizhe, Yan, Li, Ma, Zongmin
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.09.2023
• Subjects: Encoder-Decoder architecture; Graph Attention Network; Long Short-Term Memory Network; Temporal Knowledge Graph; Temporal Knowledge Graph Completion; Temporal Knowledge Graph; Encoder-Decoder architecture; Long Short-Term Memory Network; Graph Attention Network; Temporal Knowledge Graph Completion
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2023.119225

Temporal knowledge graphs (TKGs) often suffer from incompleteness, leading to an important research issue: Temporal Knowledge Graph Completion (TKGC). Knowledge Graph Embedding (KGE) methods have proven to be effective in solving this issue. However, most of them handle triples independently and do not capture complex information embedded in the neighborhood topology of central entities. To this end, we propose a Timespan-awareGraphAttention-basedEmbedding Model named T-GAE to tackle the TKGC task. To the best of our knowledge, T-GAE is the first KGE model in which Graph-Attention-Networks (GATs) and Long Short-Term Memory (LSTM) Networks are simultaneously applied to the TKGC task. In essence, our model is an Encoder-Decoder architecture, where the encoder consists of an LSTM network and a GAT network. Firstly, we employ LSTM layers to learn new time-aware relational embeddings to incorporate time information. Then, we utilize these time-aware relational embedding and GATs considered as neighborhood aggregators to learn the entity and relational features of the central entity neighborhoods. Thus, T-GAE can capture the interaction features between multi-relational facts and the abundant temporal information in TKGs. As for the decoder, we choose the ConvKB model, which is essentially a scoring function. Our experiments demonstrate that T-GAE significantly outperforms most of the existing baseline methods for TKGC in terms of MRR and Hit@1/3/10.