Can GAE work in inductive setting?

[AmoghM]

In GAE, the decoder is trying to reconstruct the adjacency matrix. I presume that this approach makes the GVAE completely transductive because an unseen node will not be present in the adjacency matrix during training and we cannot dynamically change the dimensions of the matrix either. Is there a way to make it inductive?

[tkipf]

GAE is an inductive model: the decoder takes pairs of node embeddings to predict links, whereas the encoder dynamically constructs these node embeddings from node features and their local neighborhoods -- both of these processes can be used in an inductive way on unseen parts of a graph and/or on new nodes/edges (or on a different graph with similar features/statistics altogether). If you solely use one-hot vectors as initial node features, however, any GNN model will typically not generalize to unseen parts of a graph.

On Tue, Jul 28, 2020 at 7:01 AM Amogh Mishra notifications@github.com wrote:

In GAE, the decoder is trying to reconstruct the adjacency matrix. I presume that this approach makes the GVAE completely transductive because an unseen node will not be present in the adjacency matrix during training and we cannot dynamically change the dimensions of the matrix either. Is there a way to make it inductive?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tkipf/gae/issues/65, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABYBYYBOQF7CZZTRIHZSQ7LR5ZLRNANCNFSM4PKDPMYQ .