implementation of the DropMessage paper

[rbSparky]

Recently I had planned to work on some experiments related to the DropMessage paper and its extensions and application on various datasets like

• Cora
• CiteSeer
• Pubmed
• Flickr
• ogbn-arxiv

If it would be considered useful, I could write up a tutorial/blog related to the work to show GNN.jl users how to modify architectures and use them on various datasets and experiment along the way.
Experiments include:

• Edge perturbation
• Feature noise injection
• Subgraph sampling
• Diffusion

Let me know if this will be useful for the repository! 😄
Thanks

[rbSparky]

I will use this space to add notes regarding this problem

• Core Functionality:

  • Remove edges PR Added remove_edges function #414
  • Remove nodes PR Added remove_nodes function  #420

• Dropout performs random dropping operation on the node feature matrix, while DropEdge and DropNode, as the name implies, respectively act on the adjacency matrix (edges) and nodes:

  • Drop edge PR Add drop_edge transform #419
  • Drop node PR Add drop_nodes transform #426

• Experiments

  • Edge Perturbation PR Added perturb_edges function  #423
  • PPR Diffusion PR Added Personalized PageRank Diffusion [ppr_diffusion function] #427
  • Edge Augmentation PR Added perturb_edges function  #423
  • Node Shuffle PR Added node_shuffle #429 (to be implemented by user)

[rbSparky]

After further discussion(on Slack), since it has been decided that it would be more valuable to implement more relevant applications, this discussion thread can now be used for writing the tutorial for "Molecular Property Prediction using EGNNs". It can serve as a part 2 to the current published tutorial, adding a regression task(using QM9) since currently only classification is shown
Paper

Dataset to be used: QM9

The above techniques from DropMessage can be incorporated as data augmentation technique (to show the user how to write custom functions for tasks like these)

This tutorial is in progress!