Why not use test_ mask when predicting

[Chrisable]

Why not use test_ mask when predicting？What are the benefits of using all the data ?

[rusty1s]

model.inference applies layer-wise computation of node features to prevent the neighbor explosion problem. That is, we compute all node embeddings of layer 1 before we compute node embeddings of layer 2. As such, we need to compute the node embeddings here for all nodes (applying the test_mask in a later step). If you are interested in getting the predictions for all nodes, layer-wise computation is probably the best choice to do that. If you are only interested in getting the predictions for a small set of nodes, it is not necessary to do so.

[Chrisable]

You means that if i am only interested in getting the predictions for the nodes of the test_mask,i can only use the nodes of test_mask instead of all data.If so,will the accuracy drop?I worry that if we use the nodes of test_mask, when the nodes of test_mask's neighbors are not all trained, the accuracy of aggregating its neighbors will decline during the prediction period

[rusty1s]

If you are only using test_mask nodes during inference, the model performance will likely decrease as you have access to a smaller amount of data. Note that this does not happen here since we compute the node embeddings for all nodes in all layers, not just the ones for test nodes.