This is the release code of the following paper:
Kangzheng Liu, Feng Zhao, Guandong Xu, Xianzhi Wang, and Hai Jin. Temporal Knowledge Graph Reasoning via Time-Distributed Representation Learning. ICDM 2022.
python==3.6.5
torch==1.9.0+cu102
dgl-cu102==0.8.0.post1
tqdm==4.62.3
rdflib==5.0.0
numpy==1.19.5
pandas==1.1.5
First, extract the repetitive patterns of facts under all historical-timestamp KGs and the global static KG:
python get_history_record.py --dataset YAGO
where YAGO is the name of one dataset we used in the experiment. Other datasets includes WIKI, ICEWS14, ICEWS18, ICEWS05-15, and GDELT.
Then use the following commands to train the proposed models under different datasets:
cd src
python main.py -d YAGO --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0
Finally, use the following commands to evaluate the proposed models under different datasets:
python main.py -d YAGO --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
We provide trained models for all datasets. It is noted that the commands in the
YAGO:
python main.py -d YAGO --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
WIKI:
python main.py -d WIKI --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
ICEWS14:
python main.py -d ICEWS14 --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
ICEWS18:
python main.py -d ICEWS18 --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
ICEWS05-15:
python main.py -d ICEWS05-15 --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
GDELT:
python main.py -d GDELT --train-history-len 3 --test-history-len 3 --dilate-len 1 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm --entity-prediction --gpu 0 --add-static-graph --test
Contact us with the following email address: FrankLuis@hust.edu.cn.
The source codes take RE-GCN as the backbone to implement our proposed method. Please cite both our work and RE-GCN if you would like to use our source codes.