Higher-Order Graph Convolutional Networks for Link Prediction

A PyTorch implementation of Predicting Biomedical Interactions with Higher-Order Graph Convolutional Networks (HOGCN).

Block diagram

Requirements

The codebase is implemented in Python 3.6.9 and the packages used for developments are mentioned below.

argparse                1.1
numpy                   1.19.1
torch                   1.5.0
torch_sparse            0.6.4
pandas                  1.0.1
scikit-learn            0.22.1
matplotlib              3.2.2
scipy                   1.5.0
texttable               1.6.2

Datasets

The details about dataset used in the experiments are provided in README.

Training options

Train HOGCN with the following command line arguments.

Input and output options

  --network_type      STR    Type of interaction network     Default is `DTI`.
  --fold_id           INT    Run model on generated fold     Default is 1.

Model options

  --seed              INT     Random seed.                   Default is 42.
  --epochs            INT     Number of training epochs.     Default is 50.
  --batch_size        INT     Number of samples in a batch   Default is 256.
  --early-stopping    INT     Early stopping rounds.         Default is 10.
  --learning-rate     FLOAT   Adam learning rate.            Default is 5e-4.
  --dropout           FLOAT   Dropout rate value.            Default is 0.1.
  --order             INT     Order of neighbor including 0  Default is 4.
  --dimension         INT     Dimension of each adjacency    Default is 32.
  --layers-1          LST     Layer sizes (first).           Default is [32, 32, 32, 32]. 
  --layers-2          LST     Layer sizes (second).          Default is [32, 32, 32, 32].
  --hidden1           INT     Output of bilinear layer       Default is 64.
  --hidden2           INT     Output of last linear layer    Default is 32.
  --cuda              BOOL    Run on GPR                     Default is True.
  --fastmode          BOOL    Validate every epoch           Default is True

Running HOGCN for biomedical interaction prediction

• Train HOGCN with default parameters

python3 main.py 

• Train HOGCN on DTI network with order 3 and dimension 32 for each adjacency power

python3 main.py --network_type 'DTI' --order 3 --dimension 32 

Note that order = 3 indicates P = {0, 1, 2, 3}.

Acknowledgement

The code is based on MixHop.

If you find this code useful, please cite us as:

@ARTICLE{9354550,
author={K. {Kc} and R. {Li} and F. {Cui} and A. {Haake}},
journal={IEEE/ACM Transactions on Computational Biology and Bioinformatics}, 
title={Predicting Biomedical Interactions with Higher-Order Graph Convolutional Networks}, 
year={2021},
volume={},
number={},
pages={1-1},
doi={10.1109/TCBB.2021.3059415}}