Node Classification of N Graphs

[elbasir]

Hello,

I know GCN can predict nodes in a graph but can you consider the following scenario that I have:

1- Assume there are N patients, where each patient has his/her own PPI network and I want to classify nodes "Proteins that are mutated in each patient" in each PPI network. Can GCN handle this by passing multiple graphs (Multiple PPI networks)?

2- If the answer is Yes, how to predict nodes for a new patient in the test phase as each PPI network is unique per patient in terms of mutated protein nodes VS non-mutated protein nodes?

Thanks!

[rusty1s]

1. Each patient does have a graph for which we do node classification, while we can also batch multiple graphs together. This is very similar to the PPI example we have in examples/ppi.py.
2. You would just use the learned GNN and predict nodes in the new unseen network.

[elbasir]

Thanks for your answers. I still have a couple of questions.

1- How can I validate the results per patient in the training stage?
2- How the shape of the prediction output looks like to tell about each patient's classified nodes network?

[LukeLIN-web]

1. It seems strange, why you validate the results at the training set?
2. You can print the shape print(out.shape)

[elbasir]

I am still not clear how GCN can distinguish between every validation & test split per graph if I pass multiple graphs together.

[rusty1s]

Usually, people run a round of validation after each training epoch. This is also what the examples/ppi.py script is doing.

[elbasir]

Clear, thanks a lot.