search

maciejkula / spotlight

Deep recommender models using PyTorch.
MIT License
2.99k stars 423 forks source link

besides bpr #88

Closed arita37 closed 6 years ago

arita37 commented 6 years ago

Just womdering technical hurddles for warp loss.

maciejkula commented 6 years ago

There is the adaptive hinge loss which is very similar to WARP, and consistently produces better accuracy than BPR in all my experiments. It is what I recommend using.

arita37 commented 6 years ago

Noted, thx. Have you seen the paper GAN for IR ? https://arxiv.org/abs/1705.10513

Looks like improvement of negative sampling, using more specific Proba.

On Jan 9, 2018, at 7:55, Maciej Kula notifications@github.com wrote:

There is the adaptive hinge loss which is very similar to WARP, and consistently produces better accuracy than BPR in all my experiments. It is what I recommend using.

― You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

nianjihouzz commented 6 years ago

Just curious, if the adaptive negative sampling can be generally to applied to other pair-wise loss say bpr other than hinge loss ? thanks. @maciejkula

maciejkula commented 6 years ago

@arita37 yes, I have seen it. It looks very interesting but I haven't had time to implement it.

@nianjihouzz I think that's correct, but it's not implemented.