gayanku / fdgatii

Pytorch implementation of FDGATII : Fast Dynamic Graph Attention with Initial Residual and Identity Mapping (https://arxiv.org/abs/2110.11464)

Fast Dynamic Graph Attention with Initial Residual and Identity Mapping

This repository contains a PyTorch implementation of "FDGATII : Fast Dynamic Graph Attention with Initial Residual and Identity Mapping".(https://arxiv.org/abs/2110.11464)

FDGATII combines 3 main enhancements on GAT, namely a default (safety net) initial representation is provided that is used when no similar class or node is present in the neighbourhood, so that graph attention mechanism will not fail. A more expressive universal dynamic attention is used. Finally, string regularization is applied, controlled by the hyperparameter β.

The repo, inlcuding data and datasplits used for the 10 ierations, has been forked initially from GCNII. We use the sparse (static) GATv1 attention code from pyGAT and modified dynamic attention as in GATv2.

Dependencies

• CUDA 11.3.0
• python 3.6.9
• pytorch 1.3.1

Note : FDGATII is able to run with no GPU if the GUP timing code is commented out, and then will not require CUDA.

Datasets

The data folder contains three benchmark datasets(Cora, Citeseer, Pubmed), and the new_data folder contains four datasets(Chameleon, Cornell, Texas, Wisconsin) from Geom-GCN. We use the same same full-supervised setting, and data splits, as Geom-GCN and GCNII.

Results

Testing accuracy is summarized below. We have used the 10 standard data splits ( Split 0 - 9) and obtained the Avarage Accuracy and Standard deviation.

Dataset	Depth	Dimensions	Accuracy	Std.D	Split 0	Split 1	Split 2	Split 3	Split 4	Split 5	Split 6	Split 7	Split 8	Split 9
Cora	2	64	87.7867	1.149	87.1227	89.1348	88.7324	87.7264	87.7264	86.3179	85.5131	89.1348	87.7264	88.7324
Cite	1	128	75.6434	1.8721	73.7237	75.0751	75.5255	74.9249	79.2453	73.5849	74.1742	78.979	75.6757	75.5255
Pubm	2	64	90.3524	0.297	90.5426	90.644	90.213	89.858	90.5426	90.8215	90.4412	90.2383	89.8834	90.3398
Cham	1	64	65.1754	1.8105	67.1053	66.2281	61.4035	63.5965	65.5702	64.4737	64.2544	64.693	67.9825	66.4474
Corn	1	128	82.4324	6.3095	67.5676	83.7838	91.8919	86.4865	86.4865	83.7838	83.7838	83.7838	75.6757	81.0811
Texa	2	64	82.1622	2.7562	81.0811	83.7838	81.0811	86.4865	81.0811	78.3784	83.7838	81.0811	78.3784	86.4865
Wisc	1	128	86.2745	4.4713	86.2745	80.3922	88.2353	92.1569	90.1961	82.3529	82.3529	84.3137	82.3529	94.1176

Usage

• All parameters are defined in fullSupervised_01.py.

• To run FDGATII on cora, for 1 iteration only use

python -u fullSupervised_01.py --data cora --layer 2 --alpha 0.2 --weight_decay 1e-4 --epochs 1500 --iterations 1 --mode FDGATII --support 1 --verbosity 1 --model GCNII_BASE

• To replicate the FDGATII full-supervised results, run the following script

#!/bin/bash
SCRIPT='python3.6 fullSupervised_01.py'
SETTTINGS=" --epochs 1500 --iterations 10 --mode FDGATII --verbosity 0 --model GCNII_BASE "

$SCRIPT --data cora --support 1 --layer 2 --hidden 64 --alpha 0.2 --weight_decay 1e-4 $SETTTINGS
$SCRIPT --data citeseer --support 2 --layer 1 --hidden 128  --weight_decay 5e-6 $SETTTINGS
$SCRIPT --data pubmed --support 1 --layer 2 --hidden 64  --alpha 0.1 --weight_decay 5e-6 $SETTTINGS
$SCRIPT --data chameleon --support 0 --layer 1 --hidden 64 --lamda 1.5 --alpha 0.2 --weight_decay 5e-4 $SETTTINGS
$SCRIPT --data cornell --support 2 --layer 1 --hidden  128 --lamda 1 --weight_decay 1e-3 $SETTTINGS
$SCRIPT --data texas --support 2 --layer 2 --hidden 64 --lamda 1.5 --weight_decay 1e-4 $SETTTINGS
$SCRIPT --data wisconsin --support 2 --layer 1 --hidden 128 --lamda 1 --weight_decay 5e-4 $SETTTINGS

Data sources and code

Datasets and code is forked from GCNII which uses the data set and parts of code from Geom-GCN. We use sparse attention implementation code from pyGAT as discribed in GAT . We acknowledge and thank the authors of these works for sharing their code.

Citation

@article{kulatilleke2021fdgatii,
  title={FDGATII: Fast Dynamic Graph Attention with Initial Residual and Identity Mapping},
  author={Kulatilleke, Gayan K and Portmann, Marius and Ko, Ryan and Chandra, Shekhar S},
  journal={arXiv preprint arXiv:2110.11464},
  year={2021}
}