Generative models - node attributes

Questions
#1

Hi ! I am doing some literature on graph generative models. It seems to me that most work on graph generation focuses on generating topologies i.e adjacency matrices. For instance, graph auto-encoders have P(A | Z) as output rather than P(A,X | Z) or P(X|Z)

Is there any work that covers generating new node attributes and their corresponding topology as well? One example would be generating an unseen molecular compound with a fixed topology (adjc. matrix A), and variable atom features (node feats) and bond features (edge feats). How would one proceed with such a task ?

#2

In scenarios such as molecular generation, typically both the graph topology and the particular atoms/bonds are generated simultaneously. In most cases, the node/edge features are atom/bond features that can be determined once the molecule is determined. I guess this is the reason why people rarely generate node/edge features with generative models. I’ve collected a literature collection on using graph neural networks in chemistry and biology and you may find some papers included interesting.