I want to use DropEdge to improve the DGL GSAGE accuracy on cora dataset whose link is Node Classification with DGL — DGL 1.1 documentation? Does any one tell me how can I use it? Is it beneficial to improve the accuracy for transductive node classification task?
DropEdge is generally used to enhance the robustness of GNNs to noisy edges. There are two common practices: (1) apply DropEdge to the input data as data augmentation; (2) apply DropEdge during forward to randomly mask out some edges (similar to dropout). You could refer to the DGL API here: DropEdge — DGL 1.1 documentation
1 Like
@minjie , can we use drop edge along with edge weights in forward function? Is there a way to align edge weight in forward function along with edge indexes which will be randomly removed by the dropEdge?
You can simply save edge weights with g.edata. Then when you remove the edges with the module, the corresponding edge weights will also get removed.
1 Like
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.