The result when setting L_s as non-zero?

[yangxh11]

I have noticed that in your paper, the L_s is set as 0. It is consistent with the code:
def _select_loss(self, selected_logits, labels): loss1 = 0

But I am curious about the experiment result of a non-zero L_s. Does the accuracy decrease evidently？

[chou141253]

In our experiments, using GCN combiner and L_s at the same time may cause unstable training. We would take more experiments on this, thanks.