This is the re-implementation of the GCN model, which is from the paper
SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NEURAL NETWORK by Thomas N. Kipf et al.
This implementation is a simplified version adapted from https://github.com/tkipf/gcn by removing dense and Chebyshev models.
Moreover, it supports multiple graph convolutional layers just by setting parameters of GCN object.
According to my experiments, GCN with more than two layers doesn't seem to improve the classification performance.
- python 3+
- tensorflow 1.6+
To train a default GCN model using the cora dataset:
python main.py
Optionally, you can change dataset, learning rate, the number of training epochs, dropout, weight decay and early stopping constant. For example:
python main.py --dataset citeseer --learning_rate 0.001
You can also change the number of hidden layers and their dimensionality:
python main.py --hidden_dimensions=64,64