BERT4Rec in RecBole

[zhaijianyang]

The BERT4Rec in recbole is not so bad. I ran the code and it is much better than the one in the paper. Why?

[asash]

Default parameters in the original code result in underfit model. You should train much more time than specified in the scripts.
See our reproducibility study

https://arxiv.org/abs/2207.07483

[asash]

You can also check out our hugging face based implementation https://github.com/asash/bert4rec_repro
It usually converges faster than original.

RecBole reported that they fixed some bugs in their ber4rec implementation after our repro paper, so maybe it is now not bad (thought I didn't try their new version)

[zhaijianyang]

ok thank you