Here is the news file but the Cliff Notes are:
- sub-sampling for class imbalances is now integrated with
trainand is used inside of standard resampling. There are four methods available right now: up- and down-sampling, SMOTE, and ROSE. The help page has detailed information.
- Nine additional models were added, bringing the total up to 192.
- More error traps were added for common mistakes (e.g. bad factor levels in classification).
- Various bug fixes and snake enhancements
On-deck for upcoming versions:
- An expanded interface for preprocessing. You might want to process some predictors one way and others differently (or not at all). A new interface will allow for this but should maintain backwards compatibility (I hope)
- Censored data models. Right now we are spec'ing out how this will work but the plan is to have models for predicting the outcome directly as well as models that predict survivor function probabilities. Email me (
email@example.com) if you have an interest in this.
- Enabling prediction intervals (for models that support this) using
predict.train. To be clear,
caretisn't generating these but if you fit an
lmmodel, you will be able to generate intervals from
predict.lmand pass them through