A Time-Aware Graph Attention Network for Temporal Knowledge Graphs Reasoning

Temporal Knowledge Graphs (TKGs) have been widely used in various domains to describe dynamic facts using quadruple information (subject, relation, object, timestamp). The TKG reasoning task aims to predict missing entity, i.e., "?" in the quadruple (entity, relation, ?, future time) from known facts. Although existing temporal knowledge graph reasoning models consider the quadruple information of each timestamp separately, they fail to fully exploit the temporal information as well as the hidden information in the known graph structure. To address the above issue, we propose an end-to-end encoder-decoder framework that incorporates temporal information and two-hop neighbor information into the entity embedding representation. Specifically, we design a time-aware graph attention network (TA-GAT) as an encoder. Unlike existing models that deal with each quadruple independently, we integrate two-hop neighbor nodes into TA-GAT to capture the hidden properties of the target entity domain. To further improve our model, we enhance the convolutional neural network-based knowledge graph embedding model ConvTKB as a decoder. Experimental results show that our model TA-GAT outperforms the state-of-the-art models on three datasets, i.e., WIKI, YAGO, and ICEWS14.