Customizing Sampler

[tripledoubleE]

Hi Teams, thank you for your great RecBole library!

I'm now working on the negative sampling on RecSys and using RecBole. I have some questions and I appreciate it if you give me any comments, links, and suggestions.

Recbole uses uniform distribution as a negative sampling strategy. I'm curious how the models will yield results with different distributions. Therefore, I made some changes to the sampler.py file on my local. I have some questions:

1. In the sample_by_key_ids(...) function creates (key_num * num) negative items. In the documentation, it says that key_id is candidate objective id list. When we say "candidate objective id" here, do we mean items that users do not interact with?
2. If (key_num*num) negative items are selected, aren't we selecting num negative items for each item?
3. Is sampler.py run for each user? If sampler.py runs specifically for each user, when we say dist=uniform, is it working on the entire data or only on the items that that user interacts with?
4. I want to investigate the effect of different distributions in my own project. While doing this, I want to do negative sampling more specifically for the user, taking into account the item distributions of the users separately. In other words, for each user, I want to narrow down my data set to only the items that those users interact with. Is it possible to do this using sampler.py?

I'm looking forward to hearing from you if you have time.
Thank you in advance!