Investigating the Effect of Hard Negative Sample Distribution on Contrastive Knowledge Graph Embedding

The success of the knowledge graph completion task heavily depends on the quality of the knowledge graph embeddings (KGEs), which relies on self-supervised learning and augmenting the dataset with negative triples. There is a gap in literature between the theoretical analysis of negative samples on contrastive loss and heuristic generation of quality (i.e., hard) negative triples. In this paper, we modify the InfoNCE loss to explicitly account for the negative sample distribution. We show minimizing InfoNCE loss with hard negatives maximizes the KL-divergence between the given and negative triple embedding. However, we also show that hard negatives can lead to false negatives (i.e., accidentally factual triples) and reduce downstream task performance. To address this issue, we propose a novel negative sample distribution that uses the graph structure of the knowledge graph to remove the false negative triples. We call our algorithm Hardness and Structure-aware (\textbf{HaSa}) contrastive KGE. Experiments show that our method outperforms state-of-the-art KGE methods in several metrics for WN18RR and FB15k-237 datasets.