Michael Levy
Michael Levy
https://www.rstudio.com/resources/cheatsheets/how-to-contribute-a-cheatsheet/
Errors/failures on master should go to everyone still.
E.g. capture nero-zero variance columns removed and message that they are being removed. Be sure to capture important details from the functions it calls.
- This could identify numeric and low-cardinality categoricals and calculate correlations between them - `plot.` method could call `corrplot` or `ggpairs`
From ?train: "If custom performance metrics are used (via the summaryFunction argument in trainControl, the value of metric should match one of the arguments..." - [ ] accuracy and kappa?...
[Create interactions](https://topepo.github.io/recipes/reference/step_interact.html) including polynomial terms. This shouldn't be done for tree-based models that can get at interactions other ways, but it can be very useful for regression based models. Consider...
Seems like it would be at least nice, and maybe essential, to implement this as part of the recipes pipeline so that whatever's done to training data can be applied...
I currently do this hackily inside the train lapply in tune_models. Let's us modify default random hyperparameter search parameters, customize hyperparameters, etc.
`seed` argument to `tune_models` with no default. `if (!missing(seed)) set.seed(seed); train...`.