gaoghc / dane

Deep Attributed Network Embedding

Preprocess data

Enter into the Database directory and run the corresponding script, e.g.

python script_walks_cora.py

Run experiments

Enter into the project directory and run the corresponding script, e.g.

python script_cora.py

Tips

Please create the following log folders in this project directory.

./Log/cora
./Log/citeseer
./Log/wiki
./Log/pubmed

Paper

Please cite our paper if you find the code useful for your research.

@inproceedings{gao2018deep,
  title={Deep Attributed Network Embedding.},
  author={Gao, Hongchang and Huang, Heng},
  booktitle={IJCAI},
  pages={3364--3370},
  year={2018}
}

dane's Issues

auto-encoder pretraining

Hi,

I found no description about pretraining of auto-encoder on your paper.

What benefit does it have to pretrain auto encoder? -- did you compare experiment results with/without pretraining strategy?

Thanks a lot.

error in loss functions

nn_x_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=self.w + tf.eye(tf.shape(self.neg_w)[0]), logits=pre_logit_nn_x) \ - tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(tf.diag_part(pre_logit_nn_x)), logits=tf.diag_part(pre_logit_nn_x)) nn_z_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=self.w + tf.eye(tf.shape(self.neg_w)[0]), logits=pre_logit_nn_z) \ - tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(tf.diag_part(pre_logit_nn_z)), logits=tf.diag_part(pre_logit_nn_z)) first_order_loss = tf.reduce_mean(pp_x_loss + pp_z_loss + nn_x_loss + nn_z_loss)

the labels should be self.neg_w + tf.eye(tf.shape(self.neg_w)[0])

hyperparameters

Hi,
The code is well written and absolutely reproducible.
I am facing issues with other baselines.
The paper reports that other baselines are used with default parameters.
I can't reproduce the results of wiki dataset for GAE and VGAE.
Furthermore, can you share the hyper-parameters of SAGE.