Topology-Aware Correlations Between Relations for Inductive Link Prediction in Knowledge Graphs

This repository contains the code and the datasets of Topology-Aware Correlations Between Relations for Inductive Link Prediction in Knowledge Graphs. Jiajun Chen, Huarui He, Feng Wu, Jie Wang. AAAI 2021. [arXiv]

Dependencies

The code is based on Python 3.7. You can use the following command to create a environment and enter it.

conda create --name TACT-env python=3.7
source activate TACT-env

All the required packages can be installed by running

pip install -r requirements.txt

To test the code, run the following commands.

cd code/AUC-PR
bash run_once.sh WN18RR_v1 8 1 2 0.01 0.01 demo_test 10 4

Notice that, for the first time you run the code, it would take some time to sample the subgraph.

Reproduce the Results

Usage

bash {run_once.sh | run_five.sh} <dataset>  <gamma: margin in the loss function> \
<negative_sample_size> <enclosing_subgraph_hop_number> <learning_rate> <weight_decay> \ 
<experiment_id> <max_epoch> <gpu_id> 

• { | }: Mutually exclusive items. Choose one from them.
• < >: Placeholder for which you must supply a value.

To reproduce the results, run the following commands.

#################################### AUC-PR ####################################
cd code/AUC-PR
bash run_five.sh WN18RR_v1 8 1 2 0.01 0.01 demo 10 0
bash run_five.sh WN18RR_v2 8 1 2 0.01 0.01 demo 10 0
bash run_five.sh WN18RR_v3 8 1 2 0.01 0.01 demo 10 0
bash run_five.sh WN18RR_v4 8 1 2 0.01 0.01 demo 10 0

bash run_five.sh fb237_v1 16 1 2 0.01 0.01 demo 10 0
bash run_five.sh fb237_v2 16 1 2 0.01 0.01 demo 10 0
bash run_five.sh fb237_v3 16 1 2 0.01 0.01 demo 10 0
bash run_five.sh fb237_v4 16 1 2 0.01 0.01 demo 10 0

bash run_five.sh nell_v1 10 1 2 0.01 0.01 demo 10 0
bash run_five.sh nell_v2 10 1 2 0.01 0.01 demo 10 0
bash run_five.sh nell_v3 10 1 2 0.01 0.01 demo 10 0
bash run_five.sh nell_v4 10 1 2 0.01 0.1 demo 10 0

#################################### Ranking #############################
cd code/Ranking
bash run_five.sh WN18RR_v1 8 8 2 0.01 0.01 demo 10 0
bash run_five.sh WN18RR_v2 8 8 2 0.01 0.01 demo 10 0
bash run_five.sh WN18RR_v3 8 8 2 0.01 0.01 demo 10 0
bash run_five.sh WN18RR_v4 8 8 2 0.01 0.01 demo 10 0

bash run_five.sh fb237_v1 16 8 2 0.005 0.01 demo 10 0
bash run_five.sh fb237_v2 16 8 2 0.005 0.01 demo 10 0
bash run_five.sh fb237_v3 16 8 2 0.005 0.01 demo 10 0
bash run_five.sh fb237_v4 16 8 2 0.005 0.01 demo 10 0

bash run_five.sh nell_v1 10 8 2 0.01 0.01 demo 10 0
bash run_five.sh nell_v2 10 8 2 0.01 0.01 demo 10 0
bash run_five.sh nell_v3 10 8 2 0.01 0.01 demo 10 0
bash run_five.sh nell_v4 16 8 2 0.008 0.01 demo 5 0

Remark: We run each experiment five times and report the mean results.

Citation

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

@article{DBLP:journals/corr/abs-2103-03642,
  author    = {Jiajun Chen and
               Huarui He and
               Feng Wu and
               Jie Wang},
  title     = {Topology-Aware Correlations Between Relations for Inductive Link Prediction
               in Knowledge Graphs},
  volume={35}, 
  url={https://ojs.aaai.org/index.php/AAAI/article/view/16779}, 
  number={7}, 
  journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
  year={2021}, 
  month={May}, 
  pages={6271-6278}
}

Acknowledgement

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