getting embeddings in unsupervised mode

[kuhanw]

Hello,

I noticed that the process of getting embeddings in unsupervised mode essentially involves feeding self-loops to the model, i.e. inputs1==inputs2 as seen in incremental_embed_feed_dict.

How then does one generalize to new unseen nodes? Feed all the edge connections of those nodes to the model and then average their "output1" vectors?

There is this post: https://github.com/williamleif/GraphSAGE/issues/65, the reply is not clear to me. When inferring a new node, it would not exist in the id2idx dictionary as seen in the code. In such case, it would not be possible to pass it through the model naively using the minbatch iterator. How then can we go about "aggregating its neighbors" so to speak?

Thank You,

Kuhan

[shivam1702]

Following, if ever there is a reply to this.