_neg_sample in MaskItemSequence causes infinite loop when item_set contains all items from 1 to n_items - 1."

[osmanhayek]

Hi there,

I’m having an issue with the _neg_sample function in transform.MaskItemSequence.

My dataset contains around 197,000 unique users but only 10 unique items. Unfortunately, when I set MAX_ITEM_LIST_LENGTH to 50 or 100, the code sometimes encounters a user with more than 100 interactions — and their last 100 interactions include all possible item IDs. As a result, the function ends up looping forever while trying to find a negative sample that isn’t in the user’s item set.

I also tried setting train_neg_sample_args to None, but the function still gets called during BERT4Rec training.

As a temporary fix, I modified _neg_sample so that if it fails to find a negative sample after 100 attempts, it returns [PAD] (i.e., 0).

Would this be considered an acceptable workaround for my situation?

Thanks,
Osman