How to define the loss function in unsupervised node classification and graph classification？
Could you please show me an example？
My question about DGI is that the loss function loss = dgi(features) can be used to optimize directly without considering the supervised or unsupervised learning?
Whether the only thing I need to do is revise the imported gcn model according to my task?
I think someone probably has answered you already via a different channel. I’ll just briefly explain it here for other people’s benefit. For unsupervised learning, the loss does not involves supervision signals, e.g. node labels or graph labels, while supervised learning involves that. You cannot mix them.