DURA: Duality-Induced Regularizer for Tensor Factorization Based Knowledge Graph Completion

This is the code of paper Duality-Induced Regularizer for Tensor Factorization Based Knowledge Graph Completion. Zhanqiu Zhang, Jianyu Cai, Jie Wang. NeurIPS 2020. [arXiv] [NeurIPS-Official]

Dependencies

• Python 3.6+
• PyTorch 1.0~1.7
• NumPy 1.17.2+
• tqdm 4.41.1+

Results

The results of DURA on WN18RR, FB15k-237 and YAGO3-10 are as follows.

Reproduce the Results

1. Preprocess the Datasets

To preprocess the datasets, run the following commands.

cd code
python process_datasets.py

Now, the processed datasets are in the data directory.

2. Reproduce the Results

To reproduce the results of CP, ComplEx and RESCAL with the DURA regularizer on WN18RR, FB15k237 and YAGO3-10, please run the following commands.

#################################### WN18RR ####################################
# CP
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset WN18RR --model CP --rank 2000 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 100 --regularizer DURA --reg 1e-1 --max_epochs 200 \
--valid 5 -train -id 0 -save -weight

# ComplEx
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset WN18RR --model ComplEx --rank 2000 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 100 --regularizer DURA_W --reg 1e-1 --max_epochs 50 \
--valid 5 -train -id 0 -save -weight

# RESCAL
CUDA_VISIBLE_DEVICES=2 python learn.py --dataset WN18RR --model RESCAL --rank 256 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 1024 --regularizer DURA_RESCAL --reg 1e-1 --max_epochs 200 \
--valid 5 -train -id 0 -save -weight

#################################### FB237 ####################################
# CP
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset FB237 --model CP --rank 2000 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 100 --regularizer DURA_W --reg 5e-2 --max_epochs 200 \
--valid 5 -train -id 0 -save

# ComplEx
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset FB237 --model ComplEx --rank 2000 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 100 --regularizer DURA_W --reg 5e-2 --max_epochs 200 \
--valid 5 -train -id 0 -save

# RESCAL
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset FB237 --model RESCAL --rank 512 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 512 --regularizer DURA_RESCAL --reg 5e-2 --max_epochs 200 \
--valid 5 -train -id 0 -save


#################################### YAGO3-10 ####################################
# CP
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset YAGO3-10 --model CP --rank 1000 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 1000 --regularizer DURA_W --reg 5e-3 --max_epochs 200 \
--valid 5 -train -id 0 -save -weight

# ComplEx
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset YAGO3-10 --model ComplEx --rank 1000 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 1000 --regularizer DURA_W --reg 5e-3 --max_epochs 200 \
--valid 5 -train -id 0 -save

# RESCAL
CUDA_VISIBLE_DEVICES=0 python learn.py --dataset YAGO3-10 --model RESCAL --rank 512 --optimizer Adagrad \
--learning_rate 1e-1 --batch_size 1024 --regularizer DURA_RESCAL_W --reg 5e-3 --max_epochs 200 \
--valid 5 -train -id 0 -save -weight

Citation

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

@inproceedings{NEURIPS2020_f6185f0e,
 author = {Zhang, Zhanqiu and Cai, Jianyu and Wang, Jie},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {H. Larochelle and M. Ranzato and R. Hadsell and M. F. Balcan and H. Lin},
 pages = {21604--21615},
 publisher = {Curran Associates, Inc.},
 title = {Duality-Induced Regularizer for Tensor Factorization Based Knowledge Graph Completion},
 url = {https://proceedings.neurips.cc/paper/2020/file/f6185f0ef02dcaec414a3171cd01c697-Paper.pdf},
 volume = {33},
 year = {2020}
}

Acknowledgement

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

Other Repositories

If you are interested in our work, you may find the following paper useful.

Learning Hierarchy-Aware Knowledge Graph Embeddings for Link Prediction. Zhanqiu Zhang, Jianyu Cai, Yongdong Zhang, Jie Wang. AAAI 2020. [paper] [code]