Supervised learning "only" gives you a prediction function.

But with the right tools, you'll get a lot more:

• Uncertainty quantification

• Causality

• Interpretability

• Analysis of variance

• ...

And the best news: tools in this thread work for any black box model

👇

Uncertainty quantification

Conformal prediction turns "weak" uncertainty scores into rigorous prediction intervals.

For example:

• class probabilities -> classification sets

• quantile regression -> conformalized quantile regression

https://arxiv.org/abs/2107.07511

Causality

Orthogonal/double machine learning brings causal inference to supervised learning. You can estimate treatment effects by training two models (one for treatment, one for control).

https://econml.azurewebsites.net/spec/estimation/dml.html

Interpretation

There are so many model-agnostic interpretation methods, you could write a book 😉

• SHAP, LIME for explainining individual predictions

• Permutation feature importance

• Partial dependence plots for feature effects

• ...

https://christophm.github.io/interpretable-ml-book/agnostic.html

Analysis of variance

Functions can be decomposed into lower dimensional components. Decomposition is related to interpretability, but offers advantages beyond that: An attribution of the target's variance to individual features, like ANOVA in stats.

https://christophm.github.io/interpretable-ml-book/decomposition.html

Uncertainty quantification, interpretability, ... these were usually reserved for classic statistical modeling but are now available to ML.

These model-agnostic tools give rise to a new of philosophy of modeling.

Performance-driven, yet mindful of data and model.

If you want to join me and 1000+ other mindful modelers on this journey, subscribe to my newsletter.

Let's explore the powerful mix of

• supervised learning +

• black box tools as in this thread +

• mindful statistical thinking.

https://mindfulmodeler.substack.com/