Are Message Passing Neural Networks Really Helpful for Knowledge Graph Completion?

Source code for the ACL paper Are Message Passing Neural Networks Really Helpful for Knowledge Graph Completion?

Dependencies

The key packages are the pytorch,torch-geometric/torch-scatter package. More details of the installation can be refered in https://pytorch.org/ and https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

• Python 3.7
• ordered-set==4.0.2
• torch=1.4.0+cu100
• torch-cluster=1.5.4
• torch-geometric=1.4.3
• torch-scatter=2.0.4
• torch-sparse=0.6.1
• torch-spline-conv=1.2.0

Input data

We put the data in this link: https://drive.google.com/file/d/1vQXyMP5HeM-UJBOyRDnXO28YbuZMZeLb/view?usp=sharing

Please decompress the data.zip and put the data folder in the same folder position with the run.py

• FB15k
• FB15k-237
• WN18
• WN18RR
• NELL-995

Training model:

commands to run the model:

########################  CompGCN  ######################

#### compgcn nell-995
python run.py -model 'compgcn' -read_setting 'no_negative_sampling' -neg_num 0  -score_func 'conve' -opn 'corr' -data 'NELL-995'   -lr 0.001 -batch 512  -num_workers 3 -gcn_layer 1 -hid_drop 0.3  

#### compgcn fb15k-237
python run.py -model 'compgcn' -read_setting 'no_negative_sampling' -neg_num 0  -score_func 'conve' -opn 'corr' -data 'FB15k-237'   -lr 0.001 -batch 512  -num_workers 3 -gcn_layer 1 -hid_drop 0.3  

#### compgcn wn18rr
python run.py -model 'compgcn' -read_setting 'no_negative_sampling' -neg_num 0  -score_func 'conve' -opn 'corr' -data 'WN18RR'   -lr 0.001 -batch 512  -num_workers 3 -gcn_layer 1 -hid_drop 0.3  

########################  RGCN  ######################
##rgcn nell-995
python run.py   -model 'rgcn' -read_setting 'negative_sampling' -neg_num 10  -score_func 'distmult' -data 'NELL-995' -rgcn_num_blocks 100  -lr 0.001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 2 -hid_drop 0.1 

##rgcn fb15k-237
python run.py   -model 'rgcn' -read_setting 'negative_sampling' -neg_num 10  -score_func 'distmult' -data 'FB15k-237' -rgcn_num_blocks 100  -lr 0.001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 2 -hid_drop 0.1 

##rgcn wn18rr
python run.py   -model 'rgcn' -read_setting 'negative_sampling' -neg_num 10  -score_func 'distmult' -data 'WN18RR' -rgcn_num_blocks 100  -lr 0.001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 2 -hid_drop 0.1 

##rgcn wn18
python run.py   -model 'rgcn' -read_setting 'negative_sampling' -neg_num 10  -score_func 'distmult' -data 'WN18' -rgcn_num_blocks 100  -lr 0.00001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 2 -hid_drop 0.05 

##rgcn fb15k
python run.py   -model 'rgcn' -read_setting 'negative_sampling' -neg_num 10  -score_func 'distmult' -data 'FB15k' -rgcn_num_blocks 100  -lr 0.0001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 1 -hid_drop 0.1  -init_dim 600 -gcn_dim 600 -k_w 10 -k_h 60 -no_edge_reverse

########################  KBGAT  ######################
##kbgat fb15k-237
python run.py -model 'kbgat'  -read_setting 'no_negative_sampling' -neg_num 0 -score_func 'conve' -data 'FB15k-237'  -lr 0.001 -batch 512  -l2 0. -num_workers 3  -hid_drop 0.3 

##kbgat wn18rr 
python run.py  -model 'kbgat'  -read_setting 'no_negative_sampling' -neg_num 0 -score_func 'conve' -data 'WN18RR'  -lr 0.001 -nheads 8 -batch 512  -l2 0. -num_workers 3  -hid_drop 0.3

## kbgat nell 
python  run.py  -model 'kbgat'  -read_setting 'no_negative_sampling' -neg_num 0  -score_func 'conve' -data 'NELL-995'  -lr 0.001 -nheads 8 -batch 512  -l2 0. -num_workers 3  -hid_drop 0.3  

########################  mlp  ######################
## compgcn-mlp / kbgat-mlp: nell-995
python run.py  -model 'mlp' -read_setting 'no_negative_sampling' -neg_num 0 -score_func 'conve' -data 'NELL-995'  -lr 0.001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 1 -hid_drop 0.3 

## compgcn-mlp / kbgat-mlp: fb15k237
python run.py  -model 'mlp' -read_setting 'no_negative_sampling' -neg_num 0 -score_func 'conve' -data 'FB15k-237'  -lr 0.001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 1 -hid_drop 0.3  

## compgcn-mlp / kbgat-mlp: wn18rr
python run.py  -model 'mlp' -read_setting 'no_negative_sampling' -neg_num 0 -score_func 'conve' -data 'WN18RR'  -lr 0.001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 1 -hid_drop 0.3  

####rgcn-mlp fb15k-237
python run.py -model 'mlp' -read_setting 'negative_sampling'  -neg_num 10  -score_func 'distmult' -data 'FB15k-237' -lr 0.0001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 2 -hid_drop 0.1 

####rgcn-mlp wn18
python run.py -model 'mlp' -read_setting 'negative_sampling'  -neg_num 10  -score_func 'distmult' -data 'WN18' -lr 0.0001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 2 -hid_drop 0.05 

####rgcn-mlp fb15k
python run.py -model 'mlp' -read_setting 'negative_sampling'  -neg_num 10  -score_func 'distmult' -data 'FB15k' -lr 0.0001 -batch 512  -l2 0. -num_workers 3 -gcn_layer 1 -hid_drop 0.1 -init_dim 600 -gcn_dim 600 -k_w 10 -k_h 60 

Citation:

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

@article{li2022graph,
  title={Are Message Passing Neural Networks Really Helpful for Knowledge Graph Completion?},
  author={Li, Juanhui and Shomer, Harry and Ding, Jiayuan and Wang, Yiqi and Ma, Yao and Shah, Neil and Tang, Jiliang and Yin, Dawei},
  journal={Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, {ACL} 2023, Toronto, Cananda},
  year={2023}
}

For any clarification, comments, or suggestions please create an issue or contact Juanhui.