- https://arxiv.org/abs/1907.10903
- “Over-fitting and over-smoothing are two main obstacles of developing deep Graph Convolutional Networks (GCNs)”
- At each training epoch it randomly removes edges
- the idea is that less message passing is done, so we reduce oversmoothing
- it’s similar to dropout, but it’s different
- dropout sets feature dimensions to 0, which reduces overfitting
- dropedge reduces oversmoothing AND overfitting since it removes edges in the adj matrix
- There are other techniques that the author classifies as “dropNode”, but they don’t solve the problem like dropedge does
- also graph-sparsification (the goal of removing uneeded edges) requires having an objective function to know which edges to drop.
- but edgedrop doesn’t need this