2

I have a regression problem where the output y is a single probability, i.e. real number that varies in the interval [0, 1]

While using L1 or L2 loss will very likely work well, I feel that they are not the most appropriate options considering that the range [0, 1] is already well defined.

Is Binary Cross Entropy (BCE Loss in pytorch) the most appropriate in this case?

keiv.fly
  • 1,239
  • 8
  • 14
Juan Leni
  • 1,049
  • 9
  • 24
  • 1
    Not sure what you are doing exactly, but you may have a look at beta regression https://datascience.stackexchange.com/a/57686/71442 – Peter Aug 31 '20 at 20:28
  • By L1 loss, do you mean "sum of the all the absolute differences between the true value and the predicted value" or lasso regression? – Brian Spiering Jul 19 '21 at 15:56

2 Answers2

0

At first I was going to say:

It doesn't make sense to use use cross entropy loss in a regression problem!

See explanation here.

But then I realised that if you are really trying to do regression on probabilities it could have some sense.

But still, why would you use it instead of L1, L2? So maybe try it and let me know if it works better!

pcko1
  • 3,910
  • 1
  • 14
  • 29
Francesco Pegoraro
  • 724
  • 4
  • 19
  • Well.. that is exactly what the question is about. :) Normally you would not use it, however, here it might be appropriate. I am looking for an explanation that is not based on experimentation. – Juan Leni Oct 01 '18 at 12:40
0

Predicting probabilities is can be framed as a beta regression.

That is a separate issue than adding a regularization term (i.e., L1 or L2).

Brian Spiering
  • 20,142
  • 2
  • 25
  • 102
  • I got the feeling that the OP mentioned L1 and L2 as MAE and MSE loss functions, not as regularization. – Dave Jul 19 '21 at 15:46