Swap pairwise loss for all context-aware pointwise models?

[deklanw]

This is related to #672 but a bit separate so I'm making a different discussion for it (In general pairwise losses outperform pointwise losses for top-k ranking)

I realize that these context-aware models like D&C, AutoInt, DIN, etc are mostly oriented towards CTR prediction as a binary classification task using logloss and AUC as metrics. But, isn't it very easy to swap out the loss function to be BPR or some other pairwise loss and apply them to personalized recommendation with negative sampling, etc?

I realize that the original papers implementing these methods didn't explore the ranking setting in general. But, as far as I can tell, the contributions of the papers are basically the architecture not the loss function.

Are there no context-aware models for ranking in RecBole? It's a bit strange how only ContextRecommender models prescribe an InputType

https://github.com/RUCAIBox/RecBole/blob/0.2.x/recbole/model/abstract_recommender.py#L153-L160