Inductive learning on protein-protein interactions
In GNNs, we distinguish two types of learning – transductive and inductive. They can be summarized as follows:
- In inductive learning, the GNN only sees data from the training set during training. This is the typical supervised learning setting in machine learning. In this situation, labels are used to tune the GNN’s parameters.
- In transductive learning, the GNN sees data from the training and test sets during training. However, it only learns data from the training set. In this situation, the labels are used for information diffusion.
The transductive situation should be familiar, since it is the only one we have covered so far. Indeed, you can see in the previous example that GraphSAGE makes predictions using the whole graph during training (self(batch.x, batch.edge_index)). We then mask part of these predictions to calculate the loss and train the model only using training data (criterion(out[batch...
