Gnn-based embedding methods for end2end team recommendation

[hosseinfani]

The gnn-based methods are used to learn skill embeddings for fnn/bnn feedfoward multiclass classifier (transfer-based) but we propose to use them directly for team recommendation by link prediction. This has been done in https://github.com/fani-lab/OpenGNTF but couldn't merge it to opentf mainline and reuse the test and eval logic.

We need to code it from scratch.

[hosseinfani]

#TODO: should be only for train but since we need t2v object later, we do it for other steps, even for eval!

[hosseinfani]

we make it lazy load, that is when wants to train for a fold, if the model file is there, it loads it. to re-train, the existing files should be removed or moved out of folder.

this is not the behavour for ntf-based model as they're not reused for other purposes and always retrained.

[hosseinfani]

Hi @mahdis-saeedi
The integration of e2e has now been completed. I keep this issue open in case of any bugs/enhancement during the experimentation. thanks.

[hosseinfani]

#TODO: +han or any new gnn, preferably from pyg