I got a Geometric Deep Learning problem I cannot find the solution for. I think that is an easy question for someone experienced in the subject.
I got around 1000 graphs on which I use GAT to do node classification. Every graph has 22 nodes. For every graph, only one node is labeled as 1 (the receiver).
So it looks like this:
example 1 y: [0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
example 2 y: [0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0]
example 3 y: [0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0]
example 4 y: [0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
I think I should use Cross Entropy Loss and threat it as there are 22 classes. However, I am afraid that the model will try to understand when to predict each of the classes. Instead, it should handle this problem as evaluating node per node and finally selecting the one node that has the highest chance of being the receiver. By using BCEloss I am afraid that the model does not understand that there only can be one receiver per graph.
Could someone please help me with this problem.
I am sorry if I explained it too vaguely, please ask any questions that will make it more clear. I will try to respond as fast as I can.