Michael Levy

It seems like leaving the outcome in its original form is the desired behavior, but if we want to do things with predicted classes (e.g. choosing thresholds) we likely want...

There is currently a user-customizable parameter that controls how to weight cohesion vs. ubiquity for classification (`cohesion_weight`). It currently doesn't do anything for regression. But, there is a exponent on...

From a quick check, it looks like n_levels and outcome are silently ignored, groups and id are required, and arguments passed to `pivot` are honored. The `pivot` thing is dangerous,...

Now that we've got the `all_models` flag on `evaluate`, we could have a plot method that plots a bar or dot plot for each model's performance at its best hyperparameter...

Like `missingness`. Gist is: ``` d %>% summarize_if(~ is.character(.x) | is.factor(.x) , n_distinct) ```

Let's us use it numerically or as factor

Sometimes don't want any prefix; would be nice to be able to specify `prefix = ""`

Fix panel spacing in plot.model_list

1

In 3-panel plots the bottom panel is getting squeezed. Maybe use [patchwork](https://github.com/thomasp85/patchwork) instead of cowplot.