/CAN-Pytorch

Primary LanguagePython

CAN: Co-embedding Attributed Networks

This repository contains the Python&Pytorch implementation for CAN. Further details about CAN can be found in The paper:

Zaiqiao Meng, Shangsong Liang, Hongyan Bao, Xiangliang Zhang. Co-embedding Attributed Networks. (WSDM2019)

The orignal tensorflow implementation can be found in CAN.

A semi-supervised version of the CAN model implemented by Pytorch can be found in SCAN-Pytorch.

The orignal tensorflow implementation for SCAN can be found in SCAN.

Differences with tensorflow implementation

For computing the loss directly, I move part of the optimizer.py into train.py.

There is no funtion like tf.nn.weighted_cross_entropy_with_logits() in pytorch, so I implement it by myself. To avoid the overflowing issue dropped by sigmod (when computing the torch.log(torch.sigmoid(logits)) and torch.log(1 - torch.sigmoid(logits))), I clamp the logits value from -10 to 10.

def weighted_cross_entropy_with_logits(logits, targets, pos_weight):
    logits=logits.clamp(-10,10)
    return targets * -torch.log(torch.sigmoid(logits)) *pos_weight + (1 - targets) * -torch.log(1 - torch.sigmoid(logits))

Requirements

=================

  • Pytorch (1.0 or later)
  • python 3.6/3.7
  • scikit-learn
  • scipy

Run the demo

=================

python train.py

Result

The Link prediction performance AUC&AP score :

Dataset AUC AP
BLOGCATALOG 0.820 0.822
CORA 0.989 0.988
CITESEER 0.993 0.989
DBLP 0.927 0.921
FLICKR 0.894 0.910
FACEBOOK 0.989 0.987

The Attribute inference performance AUC&AP score :

Dataset AUC AP
BLOGCATALOG 0.876 0.876
CORA 0.932 0.916
CITESEER 0.949 0.934
DBLP 0.899 0.902
FLICKR 0.856 0.844
FACEBOOK 0.976 0.973

The node classification performance Micro_F1&Macro_F1 score :

Dataset Micro_F1 Macro_F1
BLOGCATALOG 0.760 0.755
CORA 0.869 0.857
CITESEER 0.743 0.689
FLICKR 0.634 0.629