bwilder0 / clusternet Goto Github PK

May I ask how to optimize the complete graph structure (community detection)

Thanks for providing your code! Looking at models.py, it looks like the distance is computed as the dot-product between a point and a center, which is different from the negative cosine distance described in the paper? I see that the cosine distance line was commented out, which one should I use?

Hi,
Thanks for opening your source codes!
I checked the implementation of the backward passing of gradient, which is calculated using the implicit function theorem in your paper. BUT it seems that your codes directly backpropagate the loss through the iterations in cluster and doesn't replace the \partial \mu / \partial x with the partial derivative at the fix-point?

Did I miss something?

The Cora dataset consists of 2708 scientific publications classified into one of seven classes. How should the embedding layer learn properly with K set to 5, while there are 7 clusters to assign?