I've sent a slew of packages to CRAN recently (thanks to Swetlana and Uwe). There are updates to:
caretwas primarily updated to deal with an issue introduced in the last version. It is a good idea to avoid fully loading the underlying modeling packages to avoid name conflicts. We made that change in the last version and it ended up being more complex than thought. A quirk in the regression tests missed it too but the whole thing can be avoided by loading the modeling package.
recipeshad more substantial enhancments including six new steps, a better interface for creating interactions (using selectors), and the ability to save the processed data in a sparse format.
C50have been updated and brought into the age of
C5.0now has a nice plot method a la
partykitand now has a vignette. I'll be adding a few features to each over time.
Two new packages:
yardstickcontains many of the performance measurement methods in
caretbut in a format that is easier to use with
dplyrsyntax and functional programming.
tidyposterioris a Bayesian version of
resamplesfunction. It can be used to take the resampling results from multiple models and do more formal statistical comparisons. It is similar in spirit to Benavoli et al (2017). We are looking for nominations for the hex logo so please offer your suggestions (but keep it clean).
I'll be in the UK next week doing three talks in three days:
- First, I'll be giving a talk at the London R-Ladies meetup on Monday October 3rd with perhaps the best title yet: Whose Scat Is That? An 'Easily Digestible' Introduction to Predictive Modeling and
- On Tuesday, October 4th I'm giving a talk at the Cambridge RUG on tuning hyperparameters using optimization algorithms. This is an extension of this blog post.
- Finally, on Wednesday (the 5th) at the fantastic Nonclinical Statistics Conference, I'll be speaking on Statistical Mediation in Early Discovery by Bayesian Analysis and Visualization. The banner image above is from this talk. Lots of priors and shiny apps.