Question about node relationship classification with a single graph and overfitting concerns

Questions
#1

Hi everyone,

I’m working on a task to classify the types of relationships between nodes in a graph, but my issue is that I only have a single graph. I understand I need to split the graph into train, validation, and test sets, but I’m concerned about overfitting since I’m working with just one graph.

I’ve been considering using samplers to mitigate this, but I’m unsure if this would truly solve the problem. Are there any other strategies I could apply to address this and better evaluate my model?

#2

This tutorial describes how to perform edge classification training with DGL. 5.2 Edge Classification/Regression — DGL 2.3 documentation

There will be no information leakage if you make sure the test edges have never been seen during training. And I want to make sure the edges during inference are from the same single graph, right? If that’s the case, there will be no problem with this method.