Sneak peek into the new edition of Interpretable Machine Learning
It’s time to update the Interpretable Machine Learning book to the 3rd edition. This update will be both big and small. It’s big because I’m doing a lot of rewriting and a big technical move from Bookdown to Quarto. It’s small since I’m adding only two new method chapters.
I’m super excited about this update since it cleans up both the technical debt and the book itself. (well that’s more exciting for me than you). But I’m also bringing a lot of improvements to make the book more useful. For example, I’ve written dozens of times that it’s important to look at correlation before interpreting the models, but I don’t even do that myself in the current version of the book. This changes with the new edition. I’m also introducing lots of tips and warning boxes to make the book much more practical for applied projects.
I've been neglecting the book, which has to do with feeling a bit burnt out on interpretability after my PhD and the technical debt the book has accumulated, but that's a story for another post.
I’m still trying to figure out the best way to handle the synchronization of web, ebook, and physical versions. In theory, I could update the web version continuously, as I’ve done with the earlier book versions. The print version is much slower, since updates need a new ISBN and you have to check printed proofs, etc. It’s a lot of work. But the more I change the web version, the more discrepancy there is between what you read online and the physical book, and this mismatch may lead to disappointment. For this release, I’ll publish web+ebook+print all at around the same time and I’m planning the release for 1st of March.
If you already want a sneak peek into the new chapters, I have the drafts linked below. I’m looking for feedback on the chapters. Feel free to leave comments:
March 13th 2025 update: The book is now published, see https://christophm.github.io/interpretable-ml-book/
Introduction. The introduction in the current edition is such a mess. For the 3rd edition, I wanted to add a short and nice intro.
Goals of Interpretability. To make the book more practical, I have a new chapter on the goals for why you would use interpretability in the first place. This chapter was sorely needed.
Methods Overview: What the heck? The second edition didn’t have this chapter ?! Well, it kind of did, but it was all mixed up with general concepts of interpretability. Now there’s a dedicated overview.
Data and Models. The Data chapter now introduces the data in-depth and the models along with performance metrics. It also features a new classification dataset: The Palmer penguins data (very cute dataset).
Ceteris Paribus. Complex name, simple method. But don’t underestimate it, because it’s surprisingly versatile.
Leave one feature out (LOFO) Importance: Another “simple” method that works by removing a feature, retraining a model, and measuring the increase in loss. But it’s an important method to discuss, especially how it relates to permutation feature importance, conditional importance, and feature selection.
I’m grateful for any constructive feedback on the chapters.