AutoSF

The code for our paper "AutoSF: Searching Scoring Functions for Knowledge Graph Embedding" and this paper has been accepted by ICDE2020.

News: (2021.4) AutoSF-OGB for Open Graph Benchmark is released.

Readers are welcomed to fork this repository to reproduce the experiments and follow our work. Please kindly cite our paper

@inproceedings{zhang2019autosf,
  title={AutoSF: Searching Scoring Functions for Knowledge Graph Embedding},
  author={Zhang, Yongqi and Yao, Quanming and Dai, Wenyuan and Chen, Lei},
  booktitle={2020 IEEE 36th International Conference on Data Engineering (ICDE)},
  pages={433--444},
  year={2020},
  organization={IEEE}
}

Instructions

For the sake of ease, a quick instruction is given for readers to reproduce the whole process. Note that the programs are tested on Linux(Ubuntu release 16.04), Python 3.7 from Anaconda 4.5.11.

Install PyTorch (>0.4.0)

conda install pytorch -c pytorch

Get this repo

git clone https://github.com/yzhangee/AutoSF
cd AutoSF
tar -zvxf KG_Data.tar.gz 

Reproducing the searching/fine-tuning/evaluation procedure, please refer to the bash file "run.sh"

bash run.sh

Explaination of the searched SFs in the file "searched_SFs.txt":

• The first 4 values (a,b,c,d) represent h_1 * r_1 * t_a + h_2 * r_2 * t_b + h_3 * r_3 * t_c + h_4 * r_4 * t_d.

• For the others, every 4 values represent one adding block: index of r, index of h, index of t, the sign s.

You can also rely on the "evaluate.py" file to evaluate the searched SFs by manually setting the struct variable.

Related AutoML papers (ML Research group in 4Paradigm)

• Interstellar: Searching Recurrent Architecture for Knowledge Graph Embedding. NeurIPS 2020 papercode
• Efficient Neural Interaction Functions Search for Collaborative Filtering. WWW 2020 paper code
• Efficient Neural Architecture Search via Proximal Iterations. AAAI 2020. paper code
• Simple and Automated Negative Sampling for Knowledge Graph Embedding. ICDE 2019 paper code
• Taking Human out of Learning Applications: A Survey on Automated Machine Learning. Arxiv 2018 paper