Can you predict a totally unconnected datapoint with a Graph Neural Net model

Question

Let's say I am building and training a model based on Graph Neural Net to detect bot accounts in the Social Network Graph. I have a graph dataset that I will be using to train, validate and test the model. In that graph dataset, the nodes are the users and these nodes are connected to other users by certain relationships. Now when the model is trained, the model can predict in this graph if a node is a spam or not.

What about if I want to make a prediction for a random Twitter user who has no connection to this graph, can the model be able to predict it? To be more clear, I want to know can a GNN model make a prediction on data that has no relationship with the data GNN is trained upon?

Answer 1

No.

In general, you cannot use a GNN that was trained on one dataset to make predictions on some other dataset.

Having said that, there is an enormous amount of research (and ready-to-use libraries) out there, that is concerned with so-called transfer learning and domain adaptation. The idea is to take models trained on other but somehow similar data and then to adjust them with a relatively small amount of additional work to the new data. There are quite a number of "pre-trained" models out there, ready for download.