GraphAIR

Description

This is the repository for paper GraphAIR: Graph Representation Learning with Neighborhood Aggregation and Interaction. This paper has been accepted by Patten Recognition (Elsevier).

Requirements

conda create -n air python=3.6
conda activate air
pip install -r requirements.txt

Architecture

We prove that existing neighborhood aggregation scheme has difficulty in well capturing complicated non-linearity of graph data. Our work explicitly models neighborhood interaction for better capturing non-linearity of node features.

The architecture is illustrated as follows.

Usage

You can conduct node classification experiments on citation network (Cora, Citeseer or Pubmed) using the following commands:

python train.py --dataset cora --epochs 400 --w1 1.1 --w2 0.5 --w3 0.5

python train.py --dataset citeseer --epochs 3500 --w1 1.1 --w2 0.6 --w3 0.6

python train.py --dataset pubmed --epochs 400 --w1 1.1 --w2 0.9 --w3 0.6

Cite

Please cite our paper if you use this code in your own work:

@article{hu2020graphair,
  title={Graphair: Graph representation learning with neighborhood aggregation and interaction},
  author={Hu, Fenyu and Zhu, Yanqiao and Wu, Shu and Huang, Weiran and Wang, Liang and Tan, Tieniu},
  journal={Pattern Recognition},
  pages={107745},
  year={2020},
  publisher={Elsevier}
}

Acknowledgment

The structure of this code is largely based on GCN by Kipf.