GMM Conv - pseudo coordinate


I am using the “GMMConv” in the nn module However, I am having trouble understanding the pseudo coordinate. What is it exactly, and what should I set it to?


I think pseudo coordinate is common in geometric deep learning models, it describes the coordinates of nodes on a manifold. For grid graphs, you could set the pseudo coordinate of node at row i column j to (i, j). For citation networks, you could set the pseudo coordinate to something like “importance” or “reciprocal of #neighbors” of a node, depending on the data.

I think the MoNet paper has described how to set these pseudo coordinates for different tasks, and we have examples showing how to use our GMMConv API in citation network/mnist classification tasks.