Awesome

AM-GCN

Source code for KDD2020 "AM-GCN: Adaptive Multi-channel Graph Convolutional Networks"

Environment Settings

• python == 3.7
• Pytorch == 1.1.0
• Numpy == 1.16.2
• SciPy == 1.3.1
• Networkx == 2.4
• scikit-learn == 0.21.3

Usage

python main.py -d dataset -l labelrate

• dataset: including [citeseer, uai, acm, BlogCatalog, flickr, coraml], required.
• labelrate: including [20, 40, 60], required.

e.g.

python main.py -d citeseer -l 20

Data

Link

• Citeseer: Semi-Supervised Classifcation with Graph Convolutional Networks.
• UAI2010: A Unifed Weakly Supervised Framework for Community Detection and Semantic Matching.
• ACM: Heterogeneous Graph Attention Network.
• BlogCatalog,Flickr: Co-Embedding Attributed Networks.
• CoraFull: Deep Gaussian Embedding of Graphs: Unsupervised Inductive Learning via Ranking.

Usage

Please first unzip the data folders and then use. The files in folders are as follows:

citeseer/
├─citeseer.edge: edge file.  
├─citeseer.feature: feature file.  
├─citeseer.label: label file.  
├─testL/C.txt: test file. L/C, i.e., Label pre Class, L/C = 20, 40, 60.   
├─trainL/C.txt: train file. L/C, i.e., Label pre Class, L/C = 20, 40, 60.  
└─knn
   └─ck.txt: feature graph file. k = 2~9

Parameter Settings

Recorded in ./AMGCN/config/[L/C][dataset].ini
e.g. ./AMGCN/config/20citeseer.ini

• Model_setup: parameters for training AM-GCN, such as nhid1, nhid2, beta, theta...
• Data_setting: dataset setttings, such as paths for input, node numbers, feature dimensions...

Reference

@inproceedings{wang2020gcn,
  title={AM-GCN: Adaptive Multi-channel Graph Convolutional Networks},
  author={Wang, Xiao and Zhu, Meiqi and Bo, Deyu and Cui, Peng and Shi, Chuan and Pei, Jian},
  booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  pages={1243--1253},
  year={2020}
}