About the meaning of learned features about dgi HOT 2 CLOSED

Hi Fujiao,

Thanks for your issue!
I think the key part here is that the discriminator distinguishes patches in the original from ones in a corrupted graph.
Pushing those apart will have as a consequence (see the Theorems proven in our work) a likely maximisation of mutual information within the vertices of the original graph.

Hope that helps!

Thanks,
Petar

from dgi.

Hi Fujiao,

Thanks for your issue!
I think the key part here is that the discriminator distinguishes patches in the original from ones in a corrupted graph.
Pushing those apart will have as a consequence (see the Theorems proven in our work) a likely maximisation of mutual information within the vertices of the original graph.

Hope that helps!

Thanks,
Petar

Thanks for your help! :)
Best regards,
Fujiao Ji

from dgi.