>>11072308Graph CNNs are weird. I actually don't like that they use the word "convolutional" because the mathematical operations aren't actually similar to standard convolutions that you do on images.
Here's the basic gestalt:
Each node in the graph has a feature vector.
In a given time step, a node "gathers" the feature vectors of all neighbor nodes that connect to the node.
The node generates a new feature vector as a function of the old feature vector and all the neighbor vectors.
Then you repeat.
Optionally you can have feature vectors for edges. You can also specify functions that act on node/edge features during the gathering stage.
After many rounds of gathering and processing, you do you use a final layer to classify/regress on each node or each edge or the whole graph.
