FKGE — What's That AI

Differentially Private Federated Knowledge Graphs Embedding:

Data Release

The datasets we used for experiments have been partially uploaded.
To obtain all the KGs, you can find it in https://drive.google.com/file/d/1oD1Gv2RbpNzO8GWGq7SusbAmYih5r-6Q/view?usp=sharing.
Make sure to put KGs from the Google Drive into OpenKE/benchmarks.

Update: The aligned files are updated and already put in the trainse_data/aligned folder.

Package Dependencies

numpy
tensorflow 1.xx
tensorflow_probability

Baseline Embeddings

You need to run the baseline experiments to obtain the KG embeddings through the following code:

python Config.py baseline 300 100 1.0 -1

The parameters denotes mode, epoches, dimension, gan_ratio and pred_id respectively.

Note that if you want to try other embedding algorithms or some files like 1-1.txt is missing, you need to run n_n.py from OpenKE/benchmarks for each KG in /OpenKE/benchmarks/KG_1.
You can replace baseline with strategy_1 or strategy_2 to conduct the experiments with respect to FKGE.

By running baseline embeddings, you will create a experiment/ folder and the embeddings are inside experiment/0/ if you sepcify pred_id=-1.

Federated Knowledge Graphs Embedding

After obtaining KG's initital embeddings from running the baseline model (make sure there are embeddings in the experiment/0/ folder), run:

python Config.py strategy_1 300 100 1.0 0

DPFKGE

If you want to train FKGE with the PATE mechanism, in Config.py, replace

from FederalTransferLearning.hetro_AGCN_mul_dataset import GAN

with

from FederalTransferLearning.hetro_AGCN_mul_dataset_pate import GAN

Citation

Paper: https://arxiv.org/abs/2105.07615

If you use this code in your work, please kindly cite it.

@inproceedings{Peng-2021-DPFKGE,
  title={Differentially Private Federated Knowledge Graphs Embedding},
  author={Hao Peng and
          Haoran Li and
          Yangqiu Song and
          Vincent W. Zheng and
          Jianxin Li},
  booktitle={CIKM 2021},
  year={2021},
  url={https://arxiv.org/abs/2105.07615}
}

Miscellaneous

Please send any questions about the code and/or the algorithm to [email protected]

README