tkipf
commented
5 years ago Yes, this seems to be an issue that several people reported. I have seen some recent papers that could reproduce our original score (potentially by re-implementing the model themselves), but the discrepancy is most likely due to some of the recent changes made to this repository since its original release.
I recommend just reporting whichever number you get out of our implementation as a baseline score and indicate clearly that you ran the code as released in this repository. I did not find time yet to look into the issue, but the NELL dataset in the form as we used it in the original paper is of low interest in any case (it is preprocessed in a very non-standard way, and it is not clear what one would learn from doing well on this dataset). A much better fit for the NELL knowledge graph (when represented as a graph with discrete edge types) is our R-GCN model: https://arxiv.org/abs/1703.06103
On Wed, May 15, 2019 at 5:21 AM Grant notifications@github.com wrote:
Hi, tkipf , I Run on nell dataset, can't reproduce the result showed on your paper, can't get 66.0% accuracy, only around 45%, and I have watched this closed issue #14 https://github.com/tkipf/gcn/issues/14, and I use the correct hyperprameters below 'dataset' : 'nell.0.001', 'model' : 'gcn', 'learning_rate' : 0.01, 'epochs' : 200, ''hidden1' : 64, 'dropout' : 0.1, 'weight_decay' : 1e-5, 'early_stopping': 10, but got the result 45%, I used the code you released and I didn't change any of your code.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tkipf/gcn/issues/115?email_source=notifications&email_token=ABYBYYAZSZYVX6CJQUXBPKTPVN6R5A5CNFSM4HM7KNBKYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4GT2HBUA, or mute the thread https://github.com/notifications/unsubscribe-auth/ABYBYYEDDKTXREZ7TONPEPLPVN6R5ANCNFSM4HM7KNBA .
JunfengHu1996
fansariadeh
Hi, tkipf , I Run on nell dataset, can't reproduce the result showed on your paper, can't get 66.0% accuracy, only around 45%, and I have watched this closed issue https://github.com/tkipf/gcn/issues/14, and I use the correct hyperprameters below 'dataset' : 'nell.0.001', 'model' : 'gcn', 'learning_rate' : 0.01, 'epochs' : 200, ''hidden1' : 64, 'dropout' : 0.1, 'weight_decay' : 1e-5, 'early_stopping': 10, but got the result 45%, I used the code you released and I didn't change any of your code.