0
$\begingroup$

Activation function is needed in Graph Neural Network? For example, GCN is the most famous GNN. The performance of SGC, which removes the activation function from the GCN, is the same as that of the GCN.

$\endgroup$
2
  • 1
    $\begingroup$ What is SGC? Also can you give references to the performance comparison? $\endgroup$
    – gunes
    Apr 26, 2021 at 9:45
  • $\begingroup$ please see this paper: arxiv.org/pdf/1902.07153.pdf $\endgroup$ Apr 26, 2021 at 13:03

1 Answer 1

1
$\begingroup$

As the article shows, not using any nonlinearities doens't seem to hurt performance on some graph tasks. Of course, this is different from saying that a nonlinear activation function never helps, in the same way that there are (non-graph) tasks where a linear model is the best choice, and tasks where it isn't.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.