My dataset have label 0(normal) 、1（abnormal）、2（unknow）, but my model only need to learn how to predict a node is 0 or 1. So how to deal with the label 2 when I do training use gcn model in dgl? How to set train_mask when I training the model? The train_mask only have label 0 and label 1 data or also have label 2 data?
You can create a binary mask indicating if the node has a label of 0 or 1 and use it to mask nodes of label 2 in the loss computation and evaluation.
Thanks for your reply. I got it.