The code is adopted and changed according to our needs from https://github.com/tkipf/pygcn

This github provides the implementation to reproduce the results in paper `Analysis of Convolutions, Non-linearity and Depth in Graph Neural Networks using Neural Tangent Kernel' (arxiv).

• Install the required packages pip install -r requirement.txt

• cd ntk_gcn

• There are three actions possible:
  

1. To get the performance of depth=[1,2,4,8,16] on real datasets cora and citeseer, or DC-SBM using NTK, run the following script by changing the arguments accordingly
   

• Linear/ReLU GCN -- python train.py --dataset "cora" --gcn_linear 0 --gcn_skip 0 --adj_norm "row_norm" : pass --dataset as "citeseer" or "dc_sbm", --adj_norm as "col_norm" or "sym_norm" or "unnorm", for Linear GCN --gcn_linear 1

• Linear/ReLU Skip-PC or Skip-alpha -- python train.py --dataset "cora" --gcn_linear 0 --gcn_skip 1 --skip_form "gcn" --adj_norm "row_norm" : similar arguments as above and for skip alpha pass --skip_form "gcnii"

2. To get the kernels similar to the ones in the paper,

• pass --order_by_cls 1 --save_kernel 1 in the script and the kernel gets saved in the current working directory with the name 'dataset_norm_xxt_0_skip_form_depth.npy. For getting DC-SBM results pass --dataset "dc_sbm"
• The kernels can be loaded in numpy and visualized as heatmaps

3. To train the GCN of depth d,

• pass --train_gcn 1 --layers d --csigma 1 argument along with the above arguments. Note csigma should be 1 for linear GCN and 2 for ReLU GCN.