Few-shot link prediction with meta-learning for temporal knowledge graphs

With the deepening of the research on knowledge graph embedding, temporal knowledge graphs (TKGs), which are dynamic changes over time, have gradually gained the attention of researchers. Although some TKG-embedding models have been proposed, they did not perform well for certain relationships with insufficient samples, since they all require tremendous training samples. Thus, few-shot link prediction tasks, namely predicting new relation-specific quadruples by observing only a few samples, are still very challenging. In this paper, a method named meta-reasoning for TKGs (MetaRT) is proposed to solve this universal but challenging problem. MetaRT works by extracting the meta-information of a specific relation and updating it quickly, so that the model can learn the most critical information in TKG swiftly and independently. In the meantime, temporal information can be managed well by a TKG learner. Finally, through a large number of experiments, it shows that MetaRT outperforms other existing TKG-embedding models on the problem of few-shot learning.