MILP - Few-shot Inductive Link Prediction

This is the code necessary to run experiments on MILP algorithm described in the paper .

Requiremetns

dgl==0.4.2 lmdb==0.98 networkx==2.4 scikit-learn==0.22.1 torch==1.4.0 tqdm==4.43.0

Inductive relation prediction experiments

All train-graph and ind-test-graph pairs of graphs can be found in the data folder. We use WN18RR_v1 as a runninng example for illustrating the steps.

datasets

Download the datasets from https://drive.google.com/drive/folders/17akMadQa0xJm8-zgEVXACX15cTmALpON?usp=sharing

MILP

To start training a MILP model, run the following command. python train.py -d WN18RR_v1 -e WN18RR_v1

To test MILP run the following commands.

• python test_auc.py -d WN18RR_v1_ind -e WN18RR_v1
• python test_ranking.py -d WN18RR_v1_ind -e WN18RR_v1 Change the file test_auc.py while using fine-tuning mechanism The trained model and the logs are stored in experiments folder.

Citation

If you found the provided code with our paper useful in your work, we kindly request that you cite our work.

@article{yang2022few,
  title={A Few-shot Inductive Link Prediction Model in Knowledge Graphs},
  author={Yang, Ruiting and Wei, Zhongcheng and Fan, Yongjian and Zhao, Jijun},
  journal={IEEE Access},
  year={2022},
  publisher={IEEE}
}