Non-negative Sparse Knowledge Graph Embeddings
Idea
Apply Non-Negative Tensor Factorization to Knowledge Graphs to get interpretable embeddings.
Why
We want embeddings to have a few characteristics to be interpretable:
- To be explainable by a few factors -> SPARSITY
- To be efficient in the information they provide, i.e. we don't need to know that dogs don't have wheels -> NON-NEGATIVITY
Challenges/Contributions
- Not done for KGs
- Implies Tensor factorization
- There are no interpretable KGE models
- We need to find a way to improve sparsity
Code from this repository was obtained from here. I thank the authors for providing their code.