I study the code of GraphSAGE recently and notice that there are at least two layers in its implementation, i.e., input layer and output layer. (Actually, this happens in many implementations.)
If I understand right, it means at least two-hop neighbor nodes are considered in aggregation operation. So my question is: what if I just consider one-hop neighbors, I should remove the output layer manually, right?
For one-layer GraphSAGE, the input size should be the feature size and the output size should be the final output size (e.g. number of node classes). You also need to remove all activation layers.
Ok, I see. Many thanks for the clarification! 