Mastodawn

You shouldn't use over-and undersampling to "rebalance" your classes for machine learning.

Instead?

Do nothing.

Then you already have an edge over the oversamplers.

I'm ironing out the last wrinkles. Then the paperback of Interpreting Machine Learning Models With SHAP will be available 🤩.

Interested in how a feature influences machine learning model predictions on average?

Dependence plots are your friends. They plot the feature values on the x-axis and the aggregated prediction on the y-axis.

We have the following options to describe the dependence of the model predictions on a feature:

- Partial Dependence Plots
- Individual Conditional Expectation Curves
- M-Plots
- Accumulated Local Effect Plots
- SHAP Dependence plots

This testimonial made my day!

Love the framing of the Modeling Mindsets book as "conversation book" ❤️

I wrote a post titled "SHAP Is Not All You Need", because assuming that 1 method is the best for all ML interpretation contexts is a pitfall.

But if someone forced me to pick only one, I'd pick SHAP, bc it's an entire ecosystem.

That's why I also wrote:

Master machine learning interpretability with SHAP, your tool for communicating model insights and building trust in machine learning applications.

Leanpub