Graph embedding network

[ilyalasy]

Hey, would love to hear your advice. What would be the best choice of network for graph embedding that can encapsulate it's features in a "good" way? I know the question is very general, maybe I should rephrase it: What would be the criteria for choosing specific graph embedding network ?

[rusty1s]

It's hard to give a general advice on this one. However, there exists GNNs that are more suitable for a given task than others., e.g.:

• If you think your problem highly depends on the structure of the graph, you should go for more structured-focused/expressive GNNs such as GIN or PANConv.
• If you think your problem is more feature-based, e.g., which features of neighbors point to which other features, GATConv or TransformerConv is a great choice.
• If you think that "smoothing" features is sufficient for your problem, GCNConv or SAGEConv might be the way to go.
• If you have rich edge feature information, you need a GNN that can utilize those, e.g., NNConv or GMMConv.
• If you think that your problem is inherently hierarchical, you may want to try out different pooling approaches like DiffPool.

[ilyalasy]

Thanks for this guidelines, they are pretty useful!