Representing Knowledge Graph Triples through Siamese Line Graph Sampling

While many methods exist for building latent embeddings of knowledge graphs for the task of knowledge graph completion, these methods fail to generate meaningful semantic representations of the entire knowledge graph triple. To represent triples, two main architectures exist: those leveraging the line graph of the original graph to impose greater connectivity exploited via random walks, and those leveraging weak supervision signals from pre-trained embeddings in a Siamese neural architecture. While the latter approach is successful in capturing semantic relationships between similar triples, it ignores the global graph structure, instead favoring local perturbations. On the contrary, the line graph approach captures global graph information, but suffers in computational efficiency for large-scale graphs. In this work, we, for the first time, unify the two approaches by introducing a sampling strategy that reduces the dependency on the entire line graph while generating stronger supervision signals for use in Siamese neural networks. We call our approach Siamese Line Graph Sampling (SLGS). We demonstrate that this unified approach exceeds both prior approaches on two standard knowledge graph benchmarks (WN18RR and FB15K-237) with a reduction in training time of 71% and 87%, respectively.