I was recently reading the following paper: "Semi-supervised classification with Graph Convolutional Networks" by Kipf and Welling (here).

Question: When testing on datasets, why are the authors using such a low label rate?

Context: In Table 1 where the authors list the statistics for the datasets they used to benchmark the architecture, the label rates are quite small percentages (5%, 3%, <1%). I don't quite understand why this is the case...

I know this is 'semi-supervised' classification such that we have access to ALL the node features, regardless of whether they are training or testing. However, I don't know why we are using 5% label rate as opposed to, for example, 30% label rate.

Any insight would be appreciated.

  • 1
    $\begingroup$ a main reason is that it keeps the computational cost down. I’m not aware of any performance degradations by using more labels, the only thing I can think is that if you achieve similar test performance whilst training on fewer labels, this would be better as you’re less likely to be overfitting $\endgroup$
    – David
    Feb 1, 2022 at 18:05
  • $\begingroup$ @DavidIreland - many thanks for the reply! That makes sense. $\endgroup$ Feb 2, 2022 at 10:01


You must log in to answer this question.

Browse other questions tagged .