Reproducing RGCN results on WN18

aaqqxx · December 12, 2020, 2:20pm

Hi, i was trying out the code on using RGCN for link prediction https://github.com/dmlc/dgl/blob/master/examples/pytorch/rgcn/link_predict.py, by changing the dataset from FB15k-237 to WN18. However, the results i got for MRR (filtered) is around 0.69, which is significantly different from the results published in the paper (0.819). I am curious whether there is anything different in the training paradigm used?

Thank you.

classicsong · December 14, 2020, 7:24am

We are investigating this problem now.

aaqqxx · December 17, 2020, 6:21am

Thank you very much. Hope to hear from you soon

aaqqxx · December 30, 2020, 5:11pm

Hi is there any updates on this issue? Thank you!

classicsong · January 4, 2021, 4:45am

Hi, we are lacking bandwidth on this case.
BTW, did you check whether the score is Filtered-MRR or MRR?

system · February 3, 2021, 4:45am

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.