Potential new features - Githubissues

asmahani / ordinal-boost

Gradient Boosting Ordinal Regression

MIT License

0 stars 0 forks source link

Potential new features #6

Open asmahani opened 1 month ago

asmahani commented 1 month ago

[x] plotting utility (e.g. training and holdout errors vs. iteration no)
[x] cross-validation-based early stopping
[x] line search (for both g (gamma) and theta)
[x] handling of invalid thresholds (seems a rare issue)
[ ] rather than hard-coding pdf/cdf (currently normal distro, i.e., we have ordered probit), accept them as input

asmahani commented 1 month ago

Re line search, we implemented it for theta, but decided to skip it for gamma at the moment, which means we always use gamma=1.0. however, we will keep the learning rate / shrinkage parameter for the regression function, since otherwise it will rapidly overfit (i.e., when learning rate is fixed at zero).