Why is roc computed using model.z_mean even in the VGAE case? about gae

[tkipf]

Hi Daniele, thanks for your question!

I agree with you that we should have taken a sample from the learned distribution over the latents (model.z) for a fully probabilistic model at test time. The correct way to do it, is to sample and then take an expectation over the predicted links. Since this is quite expensive to evaluate, we simply take the mean of the latents in this implementation (which is slightly different from taking the expectation after predicting links). This should probably also answer your second question.

On second thought, it should be possible to approximately correct for the bias introduced with this procedure, as all the decoder does is taking an inner product and applying a sigmoid function.

If you find a significant difference in predictive scores between the two approaches, let me know!

[danielegrattarola]

The correct way to do it, is to sample and then take an expectation over the predicted links.

You mean that you'd compute adj_rec = np.dot(emb, emb.T) (line 113) multiple times with different embeddings and then take the mean of the various adj_recs?

Thanks again

[tkipf]

Yes, that’s what I meant

[danielegrattarola]

Beautiful, I get it now.

Cheers!