Final layer choice to use concat instead of pooling

[mdanb]

I noticed that in your implementation, you choose to concatenate node embeddings in the end before doing classification (in the logistic layer) instead of using pooling. Was wondering if you could elaborate on that choice.
PS I'm talking about the code here (see the comment I added):

def add_logistic_layer(self):
    logistic_layers = []
    if self.attention_head > 0:
        logistic_in_dim = [self.attention_head * self.dims[-1]]
    else:
        logistic_in_dim = [self.adjs[-1].shape[0] * self.dims[-1]] # THIS MEANS YOU'RE CONCATENATING
    for d in logistic_in_dim:
        layer = nn.Linear(d, self.out_dim)
        logistic_layers.append(layer)
    self.my_logistic_layers = nn.ModuleList(logistic_layers)

[mweiss17]

Not sure how much additional context I can give here -- perhaps @Dutil can provide a better answer.