Why haven’t I covered tidymodels?
I get this question from time to time, and with good reason. The tidymodels framework for machine learning is powerful and about as user-friendly as we can hope for in that space, bringing consistency and clarity to a process that can continue to feel ad-hoc well into the learning curve. Over time, it’s supplanted the older favorite meta-package, carat, which coincidentally (note sarcasm) was written by one the primary tidymodels architects, Max Kuhn.
I love tidymodels and use it consistently for my machine learning needs. Yet when I’ve done vids on ML techniques, I’ve almost always shied away, opting to use individual packages instead. Why?
The short answer is that I haven’t figured out how to teach it in a YouTube-friendly way yet. Vids on Equitable Equations aim to be short, self-contained, clear expositions of individual topics. While these can (and should!) be chained together to build a comprehensive whole, it’s important to me - and to the Almighty Algorithm, whom we all serve - that they stand on their own. Machine learning is hard to do this way.
One of the wonderful things about the core tidyverse is the degree to which it teaches basic data science concepts. When we learn ggplot2, we learn the grammar of graphics. When we learn tidyr, we learn about tidy data and pivoting. As we move deeper into the era of automated coding, this structure for gaining conceptual knowledge is invaluable.
But tidymodels is different. It’s a higher-order system designed to unify a large number of related tasks, some of which have very little meaning outside of a machine learning workflow.
If you’re sitting down to study logistic regression for the first time, for instance, tidymodels will simultaneously want you to consider a data split, preprocessing steps, and performance metrics, to start. This is wonderful and important, ultimately, but puts a huge burden on the learner, not to mention the YouTuber.
Over time, I’ve covered many components of the ML process in my vids, keeping them as separate and simple as possible. This has often come at the expense of data science best practices, for instance when I build models without first splitting data and conclude without computing performance metrics on an independent set. I’ve always felt that logistic regression (to name one technique) is hard enough to learn on its own without simultaneously tackling cross-validation (to name another). It hasn’t been a perfect choice, but it’s the one I’ve made, and I think my reasons are valid.
So, the TL/DR is that tidymodels is a pedagogical challenge. I’ve spent a lot of time wrestling with it, and though I’m not done yet, I hope to be soon. With any luck, you can expect a tidymodels series some time in the fall of 2026.