Add Dynamic Recurrent Basket Model (DREAM) #590
Conversation
| from tqdm.auto import trange | ||
|
|
||
|
|
||
| class Wloss(nn.modules.loss._Loss): |
There was a problem hiding this comment.
What is W loss? Can it be more informative or is it an acronym by the authors?
There was a problem hiding this comment.
@tqtg I am not quite sure about the acronym for Wloss. The current implementation is based on this implemenation https://github.com/liming-7/A-Next-Basket-Recommendation-Reality-Check/tree/main/methods/dream
The authors of the above implementation also refer the running source code to this repo https://github.com/yihong-chen/DREAM
However, these two implementations are quite different in their naming convention. I haven't read through them thoroughly. Let's take some time to check the validity of both implementations, whether they reflect the paper idea.
There was a problem hiding this comment.
Glancing at the paper, I only see BPR loss is used without additional information. I think it's better to rely on the original implementation which contains 2 variants (BPR and reordered BPR).
lthoang
changed the title
| self.loss_fct = nn.BCELoss() | ||
| self.p_loss_fct = Wloss(self.loss_uplift, 1) | ||
| self.n_loss_fct = Wloss(1, self.loss_uplift) | ||
| self.meta_loss_fct = nn.MSELoss() |
There was a problem hiding this comment.
I think meta_loss_fct is not used anywhere during training.
|
@lthoang are we actively working on this? |
Yes. I will be back to this model next week. |
Description
Related Issues
#579
Checklist:
README.md(if you are adding a new model).examples/README.md(if you are adding a new example).datasets/README.md(if you are adding a new dataset).