Code for paper Bi2E: Bidirectional Knowledge Graph Embeddings Based on Subject-Object Feature Spaces.
These experiments are based on RotatE.
Adjust the model's hyperparameters by setting arguments.py and run in run.py.
To run the code, you need the following dependency:
The results of Bi2E on WN18RR, YAGO3-10 and FB15k-237 are as follows.
| - | MR | MRR | Hits@1 | Hits@3 | Hits@10 |
|---|---|---|---|---|---|
| WN18RR | 2798 | 0.480 | 0.432 | 0.498 | 0.574 |
| YAGO3-10 | 1496 | 0.550 | 0.468 | 0.603 | 0.697 |
| FB15k-237 | 169 | 0.346 | 0.249 | 0.384 | 0.544 |
Hyper-parameters to reproduce the reuslts are set in arguments.py
Bi2E is implemented by PyTorch and runs on a NVIDIA RTX-3090. Bi2E needs different max_steps to converge on different dataset:
| dataset | WN18RR | YAGO3-10 | FB15k-237 |
|---|---|---|---|
| MAX_STEPS | 120,000 | 150,000 | 120,000 |
| Time | 2h | 4.5h | 5h |