Simple-HHEA

The code and dataset for paper Toward Practical Entity Alignment Method Design: Insights from New Highly Heterogeneous Knowledge Graph Datasets in The Web Conf 2024.

Environment

Python
Pytorch
transformers
SentencePiece
scipy
numpy
pandas
tqdm
networkx
gensim

How to Run

The model runs in 3 steps:

1. Get the name embeddings

To get the name embeddings of entities, use:

python process_name_embedding.py --data DATASET

DATASET can be icews_wiki, icews_yago or any dataset you place in the directory data.

2. Get the structure embeddings

We use Fualign to get the embeddings of entities by deepwalk. To get the structure embeddings, use:

cd fualign
python preprocess.py --l DATASET
python longterm/main.py \
	--input "data/DATASET/deepwalk.data" \
	--output "data/DATASET/longterm.vec" \
	--node2rel "data/DATASET/node2rel" \
	--q 0.7
python get_deep_emb.py --path "data/DATASET/"

DATASET is the same as the one in Step 1.

3. Run Simple-HHEA

To run Simple-HHEA, use:

python main_SimpleHHEA.py \
	--data DATASET \
	--lr 0.01 \
    --wd 0.001 \
    --gamma 1.0 \
    --epochs 1500

use --add_noise and --noise_ratio to control whether to add noise to the name embeddings and how much noise.

use --no_structure to remove structure embeddings from model.

use--no_time to remove time embeddings from model.

Or you can use:

bash run_exp.sh

to directly run Simple-HHEA on dataset icews_wiki.

How to cite

If you interested or inspired by this work, you can cite us by:

@article{jiang2023rethinking,
  title={Rethinking GNN-based Entity Alignment on Heterogeneous Knowledge Graphs: New Datasets and A New Method},
  author={Jiang, Xuhui and Xu, Chengjin and Shen, Yinghan and Su, Fenglong and Wang, Yuanzhuo and Sun, Fei and Li, Zixuan and Shen, Huawei},
  journal={arXiv preprint arXiv:2304.03468},
  year={2023}
}