Hello. While I know the differences between transductive and inductive in theory, I can't figure out what is the differences implementation between them in GNN (e.g. GCN). With GraphSage we aggregate nodes of previous hidden layer nodes with the current node. This will try to achieve us weight matrix's that could predict new nods. Transductive GCN learns the embedding and do aggregation too, but for new node it will fail to predict(?) and will have to retrain the model to predict it. My question is what is the differences in the implementation?
References I've read:
https://arxiv.org/pdf/1706.02216.pdf
https://arxiv.org/pdf/1609.02907.pdf
Thank you.