search

acbull / Unbiased_LambdaMart

Code for WWW'19 "Unbiased LambdaMART: An Unbiased Pairwise Learning-to-Rank Algorithm", which is based on LightGBM
MIT License
225 stars 50 forks source link

Why is sigma = 2? #15

Closed ffineis closed 3 years ago

ffineis commented 3 years ago

Just a question about the implementation of unbiased lambdaMART - why is the sigma coefficient = 2? Is this related to numerical stability? The paper states "sigma is a constant with default value of 2" (section 4.3), but doesn't provide a reason. Most implementations of the lambdarank gradient just defaults to sigma = 1, including LightGBM. I'm just wondering what the benefits were that drove you to pick sigma = 2 as opposed to 1.

Thank you!

acbull commented 3 years ago

Alpha indeed a hyperparameter. It influences the shape of the sigmoid and sometimes called temperature. We just choose the one that works well on our dataset.

ffineis commented 3 years ago

Ha sorry by "alpha" I meant sigma. Ok got it, thanks much.