Is dropout necessary?

[adhikarirsr]

I have noticed that dropout is used quite often in PyG examples. Is dropout used as a regularizer? If so, then why is it favored over other regularizers? Also, doesn't SGD implicitly work as a regularizer?

[rusty1s]

This depends. In general, dropout is effective as GNNs tend to heavily overfit and to be more robust to noise in the node features and edges. Batch normalization is effective, too, and there exists other variants of dropout, e.g., edge dropout via torch_geometric.utils.dropout_adj.