graph neural networks

good tutorial series: https://pytorch-geometric.readthedocs.io/en/latest/notes/colabs.html

• solves these types of problems:

  • given this social graph, is it the same class as this other social graph?
  • given the classes of a few nodes in this graph, predict the class of nodes that don’t have labels

• Why not just convert the graph into an adj matrix, and use image conv nets?

  • https://www.youtube.com/watch?v=CwsPoa7z2c8
    • here’s how we’ll construct the features as an adj matrix (for the conv net):
      • Each cell in the adj matrix are the feature vectors for edges
        • in reality, it’ll be a 3D tensor, since each edge will prob have a feature vector.
      • We can also have another matrix for the feature vectors for nodes
        • e.g. row i of the matrix represents the features for node i
    • 2 problems:
      • 1. if the graph changes (new nodes/edges), we need to retrain the entire neural net
      • 2. If we change the labels of the graph, we’ll get different adj matrices
        • 
  • So we need to make a new kind of neural network to solve graph problems

• Main problems with GNNs:

  • overfitting
  • Over-smoothing -

Graph Convolutional Network (GCN) Spectral Graph Convolutions Graph Attention Networks (GATs) (almost always outperforms GCNs) GraphSAGE graph isomorphism network (graph classification problem) Relational Graph Convolutional Networks (R-GCNs) Message Passing GNNs (MP-GNN) Graph Segment Training Graph Auto-Encoders (GAEs) Variational Graph Auto-Encoders (VGAEs) decent video explaining different schemes to incorporate edge features: https://www.youtube.com/watch?v=mdWQYYapvR8