Hello, I’m new to DGL.
I’ve run the unsupervised GraphSage sample code.
Now I want to predict an unseen node when validation, am I directly inputting the subgraph connected by an unknown node into the inference function then taking the prediction result? or I still need to input the complete graph containing an unknown node.
Thanks!
For link prediction, you only need the feature of the paired node to predict whether edge exists or not. Your setting is valid I think, but you need to ensure that the training process also follows the same setting, that how you masked out the validation part for training.
Hello,
Thanks for your reply, I`ll try it !!
My purpose is just to take the embedding result, then to do the other downstream task.
Another question is whether the inference function in the sample code is used for prediction? In this function, the neighbor sample is full sampling? Can you give me a hint about how to sample neighbors like the training when predicting?
Hi,
Actually your scenario is not a very typical scenario for unsupervised setting. The typical use case is to learn embeddings for a known graph. I would suggest you to explore more paper about recommendations such as GCMC, to handle the unseed node in your cases. I do not have certain advices for now. But feel free to add follow up questions that I might can help.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.