bsaghafi / gcnmf

PyTorch implementation of "Graph Convolutional Networks for Graphs Containing Missing Features"

GCNmf

This is a PyTorch implementation of "Graph Convolutional Networks for Graphs Containing Missing Features", Future Generation Computer Systems, 2021.

Requirements

• pytorch
• scikit-learn
• networkx
• optuna
• tqdm

Run codes

Node Classification

To run GCNmf for node classification, you can use run_node_cls.py with some options to specify dataset, missing type, missing rate, and hyper-parameters:

$ python run_node_cls.py --dataset citeseer --type struct --rate 0.1 --verbose

The following command shows the arguments of run_node_cls.py:

$ python run_node_cls.py --help

You can optimize hyper-parameters (dropout, learning rate, weight_decay) using tune_node_cls.py:

$ python tune_node_cls.py --dataset amaphoto --type struct --rate 0.1

Link Prediction

We take VGAE as the base model for link prediction and define a graph autoencoder, which employs GCNmf as an encoder. (For here, we call the link prediction model VGAEmf to distinguish the model for node classification.) To run the model for link prediction, you can use run_link_pred.py with some options:

$ python run_link_pred.py --dataset cora --type bias --rate 0.5 --verbose

You can optimize hyper-parameters (dropout, learning rate, weight_decay) using tune_link_pred.py:

$ python tune_link_pred.py --dataset citeseer --type struct --rate 0.1

Cite

@article{taguchi2021gcnmf,
  title = "Graph convolutional networks for graphs containing missing features",
  journal = "Future Generation Computer Systems",
  volume = "117",
  pages = "155 - 168",
  year = "2021",
  author = "Hibiki Taguchi and Xin Liu and Tsuyoshi Murata",
}