Readout

We use residual connections at each message passing time step t to ensure that all levels contribute to the output representation ŷ. Therefore, we define R_t as the readout at time step t and R as the overall readout function. During each time step, a neural network transforms the node’s hidden state to the desired output dimension, followed by a simple add pool to guarantee node order invariance. Finally, we use a LogSoftmax layer to yield the final output:

The final output (ŷ) is then a simple linear combination of the readouts R_t: