search

matenure / FastGCN

The sample codes for our ICLR18 paper "FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling""
519 stars 111 forks source link

The graph of FastGCN contains still the vertices of test, why the FastGCN is inductive? #21

Closed Louis-udm closed 5 years ago

Louis-udm commented 5 years ago

I understand that fastgcn performs sampling on the X in the batch, and it get a subgraph for each sampling. But the graph (the adjacency matrix A) still contains the node of the test set, why it is inductive? But usually, we should assume that when we train the model, we don’t know where the test set is. Can anyone explain it to me?

matenure commented 5 years ago

It can be either inductive or transductive. It depends on your setting. In the inductive setting, we only use adj_{train} (adj_train = adj[train_index, :][:, train_index]) for the training (as in "train_batch_multiRank_inductive_reddit_Mixlayers_sampleA.py"), instead of using the whole adjacency matrix A.

Louis-udm commented 5 years ago

ok, I got it!