yimutianyang / sigir2021-egln Goto Github PK

Hi,
Thanks for making the code public!
I have a single question about the model comparison.
In the original paper both NGCF and LightGCN, the authors mentioned that they split the data via train/valid/test. But there are no validation data in the author's Github code while your EGLN has validation data.

Did you conduct experiments on NGCF and LightGCN with your split data? or just the author's code?
I think that NGCF and LightGCN are unfair since they use the test dataset to tune hyperparameters. But LightGCN still outperforms based on your experimental results.

Thank you!

Hi,
I have one more question about your code

line 558: _hr, _ndcg = evaluate(testdata, user_items, [5,10,15,20,25,30,35,40,45,50])
line 559: valndcg[epoch] = _ndcg[10]

I could not understand why _ndcg[10] (input : testdata) is valndcg.. I think that i miss something...
please let me know

Best regards