SAGE

A tensorflow implementation of self attentive graph embedding (SAGE) in Semi-Supervised Graph Classification: A Hierarchical Graph Perspective (WWW 2019)

Details can be found in the paper:

Semi-Supervised Graph Classification: A Hierarchical Graph Perspective. Jia Li, Yu Rong, Hong Cheng, Helen Meng, Wenbing Huang, Junzhou Huang. WWW, 2019. [Paper]

Requirements

python            2.7.15
tensorflow        1.90
numpy             1.15.0
networkx          2.1
scipy             1.1.0
sklearn           0.19.1

Dataset

proteins

Model options

  --epochs                      INT     Number of epochs.                  Default is 17.
  --weight-decay                FLOAT   Weight decay of Adam.              Defatul is 5*10^-4.
  --gamma                       FLOAT   Regularization parameter.          Default is 0.19.
  --learning-rate               FLOAT   Adam learning rate.                Default is 0.01.

Example

use pretrained model

python train.py

train from scratch

python train.py --train True

Result

The average accuracy for the pretrained model is 0.80328 for proteins.