Complex Evolutional Pattern Learning for Temporal Knowledge Graph Reasoning

This is the official code release of the following paper:

Zixuan Li, Saiping Guan, Xiaolong Jin, Weihua Peng, Yajuan Lyu , Yong Zhu, Long Bai, Wei Li, Jiafeng Guo, Xueqi Cheng. Complex Evolutional Pattern Learning for Temporal Knowledge Graph Reasoning. ACL 2022.

Quick Start

Environment variables & dependencies

conda create -n cen python=3.7

conda activate cen

pip install -r requirement.txt

Download and Process data

The dataset files can be found in the project of our SIGIR 2021 paper "Temporal Knowledge Graph Reasoning Based on Evolutional Representation Learning"(RE-GCN).

First, unzip and unpack the data files,

tar -zxvf data-release.tar.gz

Offline Training with Curriculum Learing

Then the following commands can be used to train the offline models.

Pretrain models with the minimum length.

cd src
python main.py --dilate-len 1 --n-epochs 30 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm  --entity-prediction --gpu 1 -d ICEWS14s --start-history-len 3 --train-history-len 10 --test-history-len 10 --test -1  --ft_lr=0.001 --norm_weight 1

Curriculum Training.

python main.py --dilate-len 1 --n-epochs 30 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm  --entity-prediction --gpu 1 -d ICEWS14s --start-history-len 3 --train-history-len 10 --test-history-len 10 --test 0  --ft_lr=0.001 --norm_weight 1

Evaluate the offline models

To generate the evaluation results of a offline model, set the --test to 1 (for valid set) or 2 (for test set) and the --test-history-len to k (k is the optimal length of history when the MRR metric decreases in the valid set or the length is up to maximum length K) in the commands above.

For example

python main.py --dilate-len 1 --n-epochs 30 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm  --entity-prediction --gpu 1 -d ICEWS14s --start-history-len 3 --train-history-len 10 --test-history-len k --test 2  --ft_lr=0.001 --norm_weight 1

Online training data

First, train the models with timestamps in the valid set

python main.py --dilate-len 1 --n-epochs 30 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm  --entity-prediction --gpu 1 -d ICEWS14s --start-history-len 3 --train-history-len 10 --test-history-len k --test 3  --ft_lr=0.001 --norm_weight 1

Then, train the models with timestamps in the test set

python main.py --dilate-len 1 --n-epochs 30 --lr 0.001 --n-layers 2 --evaluate-every 1 --n-hidden 200 --self-loop --decoder convtranse --encoder uvrgcn --layer-norm  --entity-prediction --gpu 1 -d ICEWS14s --start-history-len 3 --train-history-len 10 --test-history-len k --test 4  --ft_lr=0.001 --norm_weight 1

Change the hyperparameters

To get the optimal result reported in the paper, change the hyperparameters and other experiment set up according to Section 5 in the paper (https://arxiv.org/pdf/2203.07782.pdf).

Citation

If you find the resource in this repository helpful, please cite

@article{li2022complex,
  title={Complex Evolutional Pattern Learning for Temporal Knowledge Graph Reasoning},
  author={Li, Zixuan and Guan, Saiping and Jin, Xiaolong and Peng, Weihua and Lyu, Yajuan and Zhu, Yong and Bai, Long and Li, Wei and Guo, Jiafeng and Cheng, Xueqi},
  journal={arXiv preprint arXiv:2203.07782},
  year={2022}
}

Davidchenw/CEN