Global graph features

[Fabien-Couthouis]

Hello,

I was wondering on how to handle general features (or context), i.e. features that are shared between all the nodes in a graph. To be clear, I am working on a Reinforcement Learning task, where my state is represented by a graph. Each node represents an entity that has features but I have also "global features" that are related to the current state (e.g. the number of steps left, current score...) and concern every node.

I was thinking of simply processing these global features in MLPs and concatenating the resulting features with all node features, but maybe it is suboptimal. Do you know a smarted way of handling this global input?

[rusty1s]

I think your suggested approach is the most prominent one. You can easily convert global features to node-wise features during training:

data.global_features = ...
data.x = torch.cat([data.x, data.global_features[data.batch]], dim=-1)

[Fabien-Couthouis]

Hi @rusty1s, thanks for your advices.