Are GNNs able to generalize to unseen edges?


I am new to DGL, and I have been working on link prediction in DGL.
I realized that link prediction dataset (provided by DGL) only involves a single graph! However, the link prediction task that I’m trying to implement can have many samples.
I have a graph where the nodes are fixed, but some links may break, and some others may appear. So the graph can have different topologies. My task is to predict the status of those edges, which is a binary classification problem. The question is: if I train the GNN model based on ONLY 1 topology (1 adjacency matrix), will it be able to predict the other topologies? Or should I include all those different topologies in my training dataset?

More clearly, are GNNs able to generalize to unseen edges?

Thank you in advance for your help.

To generalize to unseen edges (either in single graph or multiple graphs), your training set of edges must be representative enough of the entire dataset. In your case, I guess a single topology is unlikely to be representative of all topologies.

1 Like