combined messages from adjacent nodes by summing them together with the transformed current node → post-multiplying the node embedding matrix H by the adjacency matrix plus the identity A + I
Combining Current Node and Aggregated Neighbors
Residual Connections
Mean Aggregation
Sometimes it’s better to take the average of the neighbors rather than the sum; this can be superior if the embedding information is more important and the structural information less so since the magnitude of the neighborhood contributions will not depend on the number of neighbors
