Questions about GCNAE

[almafaz]

Hello again,

I have a (maybe dumb) question regarding the GCNAE model by Kipf and Welling.

In the original paper ( https://arxiv.org/abs/1611.07308), the encoder is a GCN and the decoder is a inner product.
The task in their paper is link prediction, so they reconstruct the adjacency matrix A_ by sigmod(ZZ^T) (In the non-variational version of the GAE).

If I've understood correctly, here in the domain of outlier detection, the task would be node prediction? Is that why you reconstruct the node attributes, X_, using another GCN layer? Since it wouldn't make sense to only reconstruct A?

Thank you! 😆

[kayzliu]

We follow GUIDE's paper to implement the GCNAE. This is also why we did not name it GAE.

Reconstructing node attributes or adjacency matrix depends on which type of outliers you would like to detection. As we dicussed in the benchmark paper, reconstruction of adjacency matrix helps detecting structural outlier, while reconstruction of node attributes should help detecting contextual outlier more. Furthermore, both reconstructions are used in DOMINANT, and both reconstructions can be special cases of DOMINANT by adjusting the balancing weight parameter alpha.

[almafaz]

Thank you, that makes much more sense!