HyperFormer: Enhancing Entity and Relation Interaction for Hyper-Relational Knowledge Graph Completion

This repo provides the source code & data of our paper: HyperFormer: Enhancing Entity and Relation Interaction for Hyper-Relational Knowledge Graph Completion (CIKM2023).

Dependencies

• conda create -n hyperformer python=3.7 -y
• PyTorch 1.8.1
• contiguous_params 1.0.0
• scipy 1.7.3
• tqdm 4.64.1
• fastmoe 0.2.0
  • download the fastmoe project
  • cd fastmoe folder
  • conda install "gxx_linux-64<=10" nccl -c conda-forge -y
  • pip install -e .
• If you have problems using MoE, you can directly download the one I used fastmoe.

Running the code

Dataset

• Download the datasets from Here.
• Create the root directory ./dataset and put the datasets in.
• You will get four types of datasets:
  • Mixed-percentage Mixed-qualifier: WD50K, JF17K, and Wikipeople;
  • Fixed-percentage Mixed-qualifier: WD50K_33, WD50K_66, WD50K_100, same as JF17K and Wikipeople.
  • Fixed-percentage Fixed-qualifier: WikiPeople-3, WikiPeople-4, same as JF17K.
  • Entities with Low Degree: WD50K_100_1_degree, WD50K_100_2_degree, WD50K_100_3_degree, WD50K_100_4_degree, same as JF17K and Wikipeople.

Training model

Taking the WD50K dataset as an example, you can run the following script：

sh run.sh

For other datasets, you only need to modify the following parameters, we used the same other parameters on all datasets：

• export LOG_PATH = your log path
• export SAVE_DIR_NAME = your save path
• export DATASET = the dataset you use
• export CUDA = the gpu id
• Notes: If you want to reproduce the results in Table 1, you need to set --train_mode with_valid, because all baselines use the validation set in the training process.

Notes

• When executed conda install "gxx_linux-64<=10" nccl -c conda-forge -y, if you meet the WARNING conda.core.envs_manager:register_env(50): Unable to register environment. Path not writable or missing. You should modify write permission to anaconda，e.g., sudo chown -R hzw /home/amax/anaconda3/, hzw is your username, /home/amax/anaconda3/ is anaconda path. You need see: All requested packages already installed.

Citation

If you find this code useful, please consider citing the following paper.

@article{
  author={Zhiwei Hu and Víctor Gutiérrez-Basulto and Zhiliang Xiang and and Ru Li and Jeff Z. Pan},
  title={HyperFormer: Enhancing Entity and Relation Interaction for Hyper-Relational Knowledge Graph Completion},
  publisher="32nd ACM International Conference on Information and Knowledge Management",
  year={2023}
}

Acknowledgement

We refer to the code of CoLE. Thanks for their contributions.