Skip to content

Latest commit

 

History

History
 
 

ngnn

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 

NGNN + GraphSage/GCN

Introduction

This is an example of implementing NGNN for link prediction in DGL.

We use a model-agnostic methodology, namely Network In Graph Neural Network (NGNN), which allows arbitrary GNN models to increase their model capacity.

The script in this folder experiments full-batch GCN/GraphSage (with/without NGNN) on the datasets: ogbl-ddi, ogbl-collab and ogbl-ppa.

Installation requirements

ogb>=1.3.3
torch>=1.11.0
dgl>=0.8

Experiments

We do not fix random seeds at all, and take over 10 runs for all models. All models are trained on a single V100 GPU with 16GB memory.

ogbl-ddi

performance

test set validation set #parameters
Hits@20 Hits@50 Hits@100 Hits@20 Hits@50 Hits@100
GCN+NGNN(paper) 48.22% ± 7.00% 82.56% ± 4.03% 89.48% ± 1.68% 65.95% ± 1.16% 70.24% ± 0.50% 72.54% ± 0.62% 1,487,361
GCN+NGNN(ours; 50runs) 54.83% ± 15.81% 93.15% ± 2.59% 97.05% ± 0.56% 71.21% ± 0.38% 73.55% ± 0.25% 76.24% ± 1.33%
GraphSage+NGNN(paper) 60.75% ± 4.94% 84.58% ± 1.89% 92.58% ± 0.88% 68.05% ± 0.68% 71.14% ± 0.33% 72.77% ± 0.09% 1,618,433
GraphSage+NGNN(ours; 50runs) 57.70% ± 15.23% 96.18% ± 0.94% 98.58% ± 0.17% 73.23% ± 0.40% 87.20% ± 5.29% 98.71% ± 0.22%

A 3-layer MLP is used as LinkPredictor here, while a 2-layer one is used by the NGNN paper. This is the main reason for the better performance.

Reproduction of performance

  • GCN + NGNN
python main.py --dataset ogbl-ddi --device 0 --ngnn_type input --epochs 800 --dropout 0.5 --num_layers 2 --lr 0.0025 --batch_size 16384 --runs 50
  • GraphSage + NGNN
python main.py --dataset ogbl-ddi --device 1 --ngnn_type input --use_sage --epochs 600 --dropout 0.25 --num_layers 2 --lr 0.0012 --batch_size 32768 --runs 50

ogbl-collab

Performance

test set validation set #parameters
Hits@10 Hits@50 Hits@100 Hits@10 Hits@50 Hits@100
GCN+NGNN(paper) 36.69% ± 0.82% 51.83% ± 0.50% 57.41% ± 0.22% 44.97% ± 0.97% 60.84% ± 0.63% 66.09% ± 0.30% 428,033
GCN+NGNN(ours) 39.29% ± 1.21% 53.48% ± 0.40% 58.34% ± 0.45% 48.28% ± 1.39% 62.73% ± 0.40% 67.13% ± 0.39%
GraphSage+NGNN(paper) 36.83% ± 2.56% 52.62% ± 1.04% 57.96% ± 0.56% 45.62% ± 2.56% 61.34% ± 1.05% 66.26% ± 0.44% 591,873
GraphSage+NGNN(ours) 40.30% ± 1.03% 53.59% ± 0.56% 58.75% ± 0.57% 49.85% ± 1.07% 62.81% ± 0.46% 67.33% ± 0.38%

Reproduction of performance

  • GCN + NGNN
python main.py --dataset ogbl-collab --device 2 --ngnn_type hidden --epochs 600 --dropout 0.2 --num_layers 3 --lr 0.001 --batch_size 32768 --runs 10
  • GraphSage + NGNN
python main.py --dataset ogbl-collab --device 3 --ngnn_type input --use_sage --epochs 800 --dropout 0.2 --num_layers 3 --lr 0.0005 --batch_size 32768 --runs 10

ogbl-ppa

Performance

test set validation set #parameters
Hits@10 Hits@50 Hits@100 Hits@10 Hits@50 Hits@100
GCN+NGNN(paper) 5.64% ± 0.93% 18.44% ± 1.88% 26.78% ± 0.9% 8.14% ± 0.71% 19.69% ± 0.94% 27.86% ± 0.81% 673,281
GCN+NGNN(ours) 13.07% ± 3.24% 28.55% ± 1.62% 36.83% ± 0.99% 16.36% ± 1.89% 30.56% ± 0.72% 38.34% ± 0.82% 410,113
GraphSage+NGNN(paper) 3.52% ± 1.24% 15.55% ± 1.92% 24.45% ± 2.34% 5.59% ± 0.93% 17.21% ± 0.69% 25.42% ± 0.50% 819,201
GraphSage+NGNN(ours) 11.73% ± 2.42% 29.88% ± 1.84% 40.05% ± 1.38% 14.73% ± 2.36% 31.59% ± 1.72% 40.58% ± 1.23% 556,033

The main difference between this implementation and NGNN paper is the position of NGNN (all -> input).

Reproduction of performance

  • GCN + NGNN
python main.py --dataset ogbl-ppa --device 4 --ngnn_type input --epochs 80 --dropout 0.2 --num_layers 3 --lr 0.001 --batch_size 49152 --runs 10
  • GraphSage + NGNN
python main.py --dataset ogbl-ppa --device 5 --ngnn_type input --use_sage --epochs 80 --dropout 0.2 --num_layers 3 --lr 0.001 --batch_size 49152 --runs 10

References

@article{DBLP:journals/corr/abs-2111-11638,
  author    = {Xiang Song and
               Runjie Ma and
               Jiahang Li and
               Muhan Zhang and
               David Paul Wipf},
  title     = {Network In Graph Neural Network},
  journal   = {CoRR},
  volume    = {abs/2111.11638},
  year      = {2021},
  url       = {https://arxiv.org/abs/2111.11638},
  eprinttype = {arXiv},
  eprint    = {2111.11638},
  timestamp = {Fri, 26 Nov 2021 13:48:43 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2111-11638.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}