Why cant we use negative sampling for sequential recommender?

[Jeriousman]

In Recbole now, It seems that sequential recommenders are not utilizing BPR loss with negative sampling? Why is that the case?

Generating negative sequence shouldnt help? Even BERT4Rec official paper noted that they did negative sampling but BERT4Rec in Recbole doesnt seem to support negative sampling as well.

Can someone please explain why Recbole isnt doing so?

[AoiDragon]

Hello @Jeriousman , sequential recommenders in RecBole support BPR loss with negative sampling. You can set loss_type to BPR and specify train_neg_sample_args. For more detailed information, you can refer to our API doc. Here is an example of GRU4Rec.