HIN-Align

Bachelor degree research on heterogeneous information network alignment. Extended work of GCN-Align.

Datasets are from JAPE and IONE.

Environment

• python>=3.5
• tensorflow>=1.10.1
• scipy>=1.1.0
• networkx>=2.2

Test

unzip dbp15k.zip
chmod +x test.sh
./test.sh

The pre-trained results are in the res/ folder. If you don't want to train by yourself, just see the files in it.

For social network, run:

python train_sn.py --seed 5

For automatically train weights:

python train_auto.py

Citation

Please politely cite our work as follows:

Zhichun Wang, Qingsong Lv, Xiaohan Lan, Yu Zhang. Cross-lingual Knowledge Graph Alignment via Graph Convolutional Networks. In: EMNLP 2018.

TODO

• [x] Change a_ij to sigmoid(a_ij)
• [x] Combine with TransE (KG) or DeepWalk (SN)
• [ ] Combine with MT
• [x] Social Network Alignment
• [ ] Iterative or Bootstrapping
• [ ] Use faiss to improve evaluation speed
• [ ] Dimension Reduction or other ways of combination
• [x] Automatic training for hybrid weights
• [ ] Batched training for GCN
• [ ] Try other GNN models