Interpreting how does RotatE handle 1-N, N-N type relations...

DeepGraphLearning / KnowledgeGraphEmbedding

MIT License

1.24k stars 264 forks source link

Interpreting how does RotatE handle 1-N, N-N type relations... #36

Closed agoel00 closed 4 years ago

agoel00 commented 4 years ago

Hi. Great work and repo. I was curious to know what is the geometrical interpretation when RotatE handles 1-N and N-N type relations. For example, in TransH we know that the embeddings are projected onto a relation specific hyperplane to bypass this issue.

How is this handled in the complex plane of RotatE?

Edward-Sun commented 4 years ago

Hi, thanks for the good question. In RotatE, we don't consider 1-N, N-N type relations. This is a problem worth studying as future work.