Which strategy is used for negative sampling??

[tripledoubleE]

Hello,
I have a question about negative sampling. In Recbole documentation, it says: "we will set positive label for all the observed interactions, and select some negative samples from the unobserved interactions according to a certain strategy." Can you explain the strategy this library uses while implementing the negative sapling for labeling the data?

Thanks in advance.

[Sherry-XLL]

Hello @tripledoubleE, thanks for your attention to RecBole!

In RecBole, we support three commonly used negative sampling methods, including random negative sampling (RNS), popularity-biased negative sampling (PNS) and dynamic negative sampling (DNS).

Training Settings — RecBole 1.1.1 documentation

RecBole samples num negative items for each user in a batch. We first implement get_used_ids to obtain used_ids, which are composed of positive items. When negative sampling, we will not sample from these items. Then we implement self.sampling to sample negative items. If the distribution is uniform, the negative items will be sampled by np.random.randint in a uniform distribution. Else if the distribution is popularity, the sample distribution will be based on item popularity. And, dynamic (bool) decides whether we adopt dynamic negative sampling.

If you have any usage issues with our library, we suggest that you open a new issue instead of a discussion, so that we can better answer your questions.

[tripledoubleE]

thank you so much