wehlutyk
commented
6 years ago Recent thoughts about this:
- a visual comparison already says an awful lot: if the reconstructed adjacency and feature matrices look like the original, we're on the right path
- for features (and for adjacency), we can use Adversarial Attacks on Node Embeddings. This is also a great metric to try and beat: existing algorithms are affected X% by this attack, and how much better do we do by taking both features and structure into account?
For adjacency, this should be the standard link prediction metric (look it up in one of the reference papers).
I'm not sure what we should use for feature reconstruction yet.