Enhance Temporal Knowledge Graph Completion via Time-Aware Attention Graph Convolutional Network.

HaoHui Wei,Hong Huang,Teng Zhang,Xuanhua Shi,Hai Jin

DOI: https://doi.org/10.1007/978-3-031-26390-3_8

2022-01-01

Abstract:Previous works on knowledge graph representation learning focus on static knowledge graph and get fully developed. However, task on temporal knowledge graph is far from consummation because of its late start. Recent researches have shifted to the temporal knowledge graph relying on the extension of static ones. Most of these methods seek approaches to incorporate temporal information but neglect the potential adjacent impact merged in temporal knowledge graphs. Meanwhile, different temporal information of involved facts evoke impact with different extent on the concerned entity, which is always overlooked in the previous works. In our paper, we propose a Time-aware Attention Graph Convolutional Network, named TAGCN, for temporal knowledge graph completion. Entity completion can be turned into interactions between entity and associated neighborhood. We utilize a graph convolutional network with a novel temporal attention layer to obtain neighboring information at all timestamps to avoid diachronic sparsity. We conduct extensive experiments on various datasets to evaluate our model performance. The results illustrate that our model outperforms the state-of-the-art baselines on entity prediction.