Open KID-22 opened 1 year ago
This is good job. However, we always use BPRLoss rather than HingeLoss in pairwise learning to rank since the margin of HingeLoss is hard to tune. So I wonder whther you have tried the BPRLoss?
No, I haven't tried it. I found set margin=0 in margin loss run very well. I think other losses can achieve similar results (but you should try it).
This is good job. However, we always use BPRLoss rather than HingeLoss in pairwise learning to rank since the margin of HingeLoss is hard to tune. So I wonder whther you have tried the BPRLoss?