Is this phenomenon normal?

[Chrisable]

When I delete 20% of the training nodes in the CORA dataset, I find that the accuracy is 1% higher than when I didn't delete it, and only when I delete more than 50% of the training nodes, the accuracy will drop significantly.Why is this happening? it is because removing some nodes makes the model more robust?

[rusty1s]

That's a cool finding. Does that also hold true when training the models multiple times (e.g., 100 times) and compare the mean accuracy over the models?

[Chrisable]

I used your GCN example to train CORA, but used the subgraph to randomly remove some nodes before training

[Chrisable]

Everything else is consistent

[rusty1s]

Yeah, can you share the code you used to generate subgraphs?

[Chrisable]

Yes, please ask what is your email address?

[rusty1s]

You can look it up on my GitHub profile. Looking forward to hear from you.