Sigmoid function in DOMINANT?

[almafaz]

Hi again!

I was reading the papers and comparing their architectures to your implementation (which is awesome!) and I noticed that both the AnomalyDAE and DOMINANT papers uses the sigmoid(ZZ^T) method to reconstruct the adjacency matrix.

In your AnomalyDAE implementation you indeed use the sigmoid function to compute the matrix (line 398) but for DOMINANT (line 309), where do you apply the sigmoid function?

Thank you!

[kayzliu]

In the structural decoder, there is an activation function, which can be sigmoid function.

[lumenWAHX]

Hello, I am also reading papers on DOMINANT and have the same question.

For the activation function in the structural decoder, isn't it applied to Z? So the structural reconstruction would be sigmoid(Z)sigmoid(Z)^T. But I think sigmoid function should be applied to (ZZ^T)?

Thank you!

[kayzliu]

Gotcha, I understand what you're saying now. Our implementation does not have a sigmoid applied to ZZ^T. However, I try to add the sigmoid just now, and I test it on Cora. It even hamper the performance of detecting structural outliers. If you have different empirical results, please let me know.

[lumenWAHX]

Thank you for answering!
I too tried adding the sigmoid function to DOMINANT and applying it to Cora. It could not detect structural outliers as well, which confuses me, because in the paper Deep Anomaly Detection on Attributed Networks, the authors specify there to be a sigmoid in the structural reconstruction. I think theoretically the sigmoid function ensures that each entry of the reconstructed adjacency matrix is between 0 and 1. If there is no sigmoid function, how do we ensure the reconstruction is between the 0 and 1 range?

[jweihe]

Thank you for answering! I too tried adding the sigmoid function to DOMINANT and applying it to Cora. It could not detect structural outliers as well, which confuses me, because in the paper Deep Anomaly Detection on Attributed Networks, the authors specify there to be a sigmoid in the structural reconstruction. I think theoretically the sigmoid function ensures that each entry of the reconstructed adjacency matrix is between 0 and 1. If there is no sigmoid function, how do we ensure the reconstruction is between the 0 and 1 range?

I have same question，have you solved it？