Here is a list of topics that would be great to add:

1. Graph Neural Networks.
2. Bayesian Statistics, Probabilistic programming and application for A/B testing.
3. DBSCAN in the Clustering section.
4. SVD, Factorization Machines, and Neural Network approaches in RecSys section
5. Deep Reinforcement learning

1. In the CLI, you can add a requirement for a variance limit and math. the expectation is bounded. The Cauchy distribution is sometimes encountered in engineering in the laser industry.

2. Bias - an error with the choice of a set of functions from which you choose, the objective function is far from the projection on this set of functions that you are considering.

3. Variance - associated with the instability of the solutions obtained based on sampling examples from the population.
   For (2) and (3) better read the Statistical Learning book, the definitions from CheatSheet are too vague at least for me.

4. In non-parametric models, these are models where there are no training parameters. For example, the nearest neighbor method. (Or more sophisticated models from MetaLearning that use in one of it's flavor non-parametric blocks)

5. Cross-Validation - does not do any validation, rather it is an assessment. If you want validation (in the sense of guarantees) - this principle is not suitable.

6. In Subset selection you put 0-norm, that's very bad jargon. Zero is not the norm. It is not homogeneous (multiplying a vector by 10 will not multiply a value by 10). Use the word cardinality(.)

In the CART model, splitting occurs from the beginning of the entire tree, so CP has a more complex meaning.

Specifically, in the CART model, this is a penalty in the objective function of the form "cp|T|", where T is the number of tree nodes. Technically, CP is still included as a split score, but there is a trick - splitting the tree (what for classification, what for regression is done in general over the entire depth).

Further, in the classification mode, CART uses Gini Index, C4.5 uses Entropy.

I was a student of J. Friedman and he told me about it.
If you want to learn more about CART, I have a couple of notes
https://sites.google.com/site/burlachenkok/articles/decision-trees-parti-decision-trees-for-regression,
https://sites.google .com/site/burlachenkok/articles/some-problems-with-decision-tree

and two talks at MIPT (but they are in Russian)
https://www.youtube.com/watch?v=r4ZTy90233w
https://www.youtube.com/ watch?v=evkzN6AZTnc&t=53s

First page. "Variable Importance - ranks variables by their ability to
minimize error when split upon, averaged across all trees".

Estimating the importance of a variable is quite an abstract and broad thing.

You can do this in a large number of ways if you require theory (and generally infinite - if you don’t require theory).

Things I know about evaluation variables are here https://sites.google.com/site/burlachenkok/some-ways-to-intepretate-black-box-models

The text from this cheat sheet in the context of decision trees was suggested by J.Friedman's friend - L. Breiman from 1983, J. Friedman has a theorem from 1999 that this is in some sense a good thing.

I suggest to add extra text, because one more time think that Variable importance can be done only in one way - is deeply wrong.

And thanks for your work