SE-GNN

This is the PyTorch implementation of the Semantic Evidence aware Graph Neural Network (SE-GNN) for Knowledge Graph Embedding task, as described in our paper:

Ren Li, Yanan Cao, Qiannan Zhu, Guanqun Bi, Fang Fang, Yi Liu, Qian Li, How Does Knowledge Graph Embedding Extrapolate to Unseen Data: a Semantic Evidence View (AAAI'22)

Dependencies

• PyTorch >= 1.9.1
• DGL >= 0.7.2 (for graph neural network implementation)
• Hydra >= 1.1.1 (for project configuration)

Usage

• code/: includes code scripts.
• data/:
  • dataset/FB15k_237/: FB15k-237 dataset resources
  • dataset/WN18RR/: WN18RR dataset resources
  • output/FB15k_237/: model outputs for FB15k-237 dataset
  • output/WN18RR/: model outputs for WN18RR dataset
• config/: We use Hydra toolkit to manage the model hyper-parameters, which can be either stored in YAML file or passed by command-line. More details can be seen in official docs.
  • config.yaml: project configurations and general parameters
  • dataset/FB15k_237.yaml: best hyper-parameters for FB15k-237 dataset
  • dataset/WN18RR.yaml: best hyper-parameters for WN18RR dataset
• drawing/: includes materials to re-draw the paper figs.
  • data_analyse.py: drawing code
  • FB15k_237/ and WN18RR/: includes computed Semantic Evidence metrics data2metrics and reproduced baseline prediction xx_rank. More details can be found in our paper.

Model Training

# enter the project directory
cd SE-GNN-main

# set the config/config.yaml `project_dir` field to your own project path

# FB15k-237
python code/run.py

# WN18RR
python code/run.py dataset=WN18RR

# We store the best hyper-parameters in dataset's corresponding .yaml file, 
# and there are two ways to re-set the parameter:
# 1. Directly modify the .yaml file;
# 2. Pass the value by command-line, like:
# python code/run.py dataset=WN18RR rel_drop=0.2 ...

# draw the FB15k-237 pictures (paper Figure 2)
python drawing/data_analyse.py

# draw the WN18RR pictures (paper Figure 6)
python drawing/data_analyse.py dataset=WN18RR

The model takes about 10h for training on a single GPU, and the GPU memory cost is about 11GB for FB15k-237 and 3GB for WN18RR dataset.

Citation

Please cite the following paper if you use this code in your work:

@inproceedings{li2022segnn,
    title={How Does Knowledge Graph Embedding Extrapolate to Unseen Data: a Semantic Evidence View},
    author={Ren Li and Yanan Cao and Qiannan Zhu and Guanqun Bi and Fang Fang and Yi Liu and Qian Li},
    booktitle={Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2022},
    year={2022}
}

If you have any questions about the paper or the code, please feel free to create an issue or contact Ren Li <[email protected]>. Sometimes I may not reply very quickly because of the engaged matters, but I will do it asap when I am free :)