search

Tiiiger / SGC

official implementation for the paper "Simplifying Graph Convolutional Networks"
MIT License
831 stars 145 forks source link

20NG Training Set Reported Results Question #34

Closed jharrang closed 2 years ago

jharrang commented 2 years ago

For your 20NG results reported in your original paper (88.5 ± 0.1), was the model trained on the full public 20NG train set, of size 11314, or were the reported results generated using the code currently in this repo, which appears to exclude the validation set from the data used to train the model? (The latter would use a training set of size 10183).

Also, I notice that in Table 4 you reported a different result (87.9 ± 0.2) for the cited Yao GCN paper than they did themselves (0.8634 ± 0.0009). Do you know why that is the case?

Thanks!

Tiiiger commented 2 years ago

hi @jharrang

sorry for a late reply.

hoestly it has been a while that I don't remember the exact details. But I am sure that the code we used in this repo is the code we used to get the reported results. Historically, the data pipeline is taken from this repo (https://github.com/codeKgu/Text-GCN).

and sorry I am not sure why did we get better results for their method ...