Skip to content

Latest commit

 

History

History
 
 

compGCN

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
data_loader.py
data_loader.py
 
 
get_fb15k-237.sh
get_fb15k-237.sh
 
 
get_wn18rr.sh
get_wn18rr.sh
 
 
main.py
main.py
 
 
models.py
models.py
 
 
utils.py
utils.py
 
 

README.md

DGL Implementation of the CompGCN Paper

This DGL example implements the GNN model proposed in the paper CompositionGCN. The author's codes of implementation is in here

Example implementor

This example was implemented by zhjwy9343 and KounianhuaDu at the AWS Shanghai AI Lab.

Dependencies

  • pytorch 1.7.1
  • dgl 0.6.0
  • numpy 1.19.4
  • ordered_set 4.0.2

Dataset

The datasets used for link predictions are FB15k-237 constructed from Freebase and WN18RR constructed from WordNet. The statistics are summarized as followings:

FB15k-237

  • Nodes: 14541
  • Relation types: 237
  • Reversed relation types: 237
  • Train: 272115
  • Valid: 17535
  • Test: 20466

WN18RR

  • Nodes: 40943
  • Relation types: 11
  • Reversed relation types: 11
  • Train: 86835
  • Valid: 3034
  • Test: 3134

How to run

First to get the data, one can run

sh get_fb15k-237.sh
sh get_wn18rr.sh

Then for FB15k-237, run

python main.py --score_func conve --opn ccorr --gpu 0 --data FB15k-237

For WN18RR, run

python main.py --score_func conve --opn ccorr --gpu 0 --data wn18rr

Performance

Link Prediction Results

Dataset FB15k-237 WN18RR
Metric Paper / ours (dgl) Paper / ours (dgl)
MRR 0.355 / 0.349 0.479 / 0.471
MR 197 / 208 3533 / 3550
Hit@10 0.535 / 0.526 0.546 / 0.532
Hit@3 0.390 / 0.381 0.494 / 0.480
Hit@1 0.264 / 0.260 0.443 / 0.438