Message Passing GNNs (MP-GNN)

• https://wandb.ai/graph-neural-networks/spatial/reports/An-Introduction-to-Message-Passing-Graph-Neural-Networks—VmlldzoyMDI2NTg2

  • they are graph neural networks but you are aggregating the features from nearby nodes. that’s it!

• How are MP-GNNs different from Graph Convolutional Network (GCN)

  • People refer to them as the same things.
    • This article does say that the difference is: message passing is a way to approximate graph convolutions.
      • Message passing is like taking into account the one-hop embeddings of your neighbouring nodes.
        • This is what one layer does in a GNN
      • if we stack n layers together, we’re doing message passing n times
        • this is approximating a graph convolution that takes in features from n hops away
        • an analogy: in image convolutions, a 3x3 kernel takes in features from yourself and your neighbours. but we can also use a 5x5 kernel, which takes in even more features from further
        • stacking more layers is like increasing your kernel size

• main problems:

  • Since MPNNs are limited by problems of over-smoothing, over-squashing, and low expressivity against the WL test [1, 54], these layers could irreparably fail to keep some information in the early stage
    • https://openreview.net/pdf?id=lMMaNf6oxKM