txt A Tabulated Compendium of Data Science & Machine Learning: A Complete Reference

Preface
This book began as a private habit. For years I kept a set of notebooks — first paper, then plain text, then something more organised — in which I wrote down the things I kept having to look up. The exact form of the Cramér–Rao bound. Whether LightGBM grows leaf-wise or level-wise. What the second argument to the Benjamini–Hochberg procedure actually controls. Which of the seventeen ways of measuring a forecast error is safe when the series passes through zero. The notebooks grew, and at some point they stopped being notes and became a reference. This is that reference, rewritten for other people.
The field has a peculiar shape. It is enormous horizontally and shallow in most places: a working practitioner touches linear algebra, hypothesis testing, database formats, gradient boosting hyperparameters, transformer attention variants, Kubernetes, drift detection, and the European regulatory calendar, often in the same week, and needs about two pages of each. But in a handful of places it is very deep indeed, and shallow knowledge is dangerous there — in leakage, in causal claims, in calibration, in the difference between a model that scores well and a model that works. A reference has to serve both needs. So this one is built in two registers.
Tables carry the facts that have a shape: algorithm comparisons, hyperparameter defaults, complexity classes, metric definitions, library equivalents, hardware specifications, test assumptions. When a thing has columns, it is a table. There are over two hundred of them, and they are the part of the book you will use most often once you have read it once.
Prose carries the things that do not have a shape: why a method exists, what it assumes, how it fails, what its practitioners argue about, and what the table cannot tell you. Every table in this book is preceded or followed by prose that says what to actually do with it. A table of clustering validity indices without the paragraph explaining that all of them are internal criteria that reward the geometry you happened to assume is worse than useless.
What this book assumes
It assumes you can read a formula without being frightened, that you have written some code, and that you know roughly what a model is. It does not assume a mathematics degree, a machine learning degree, or familiarity with any particular framework. Where a derivation matters — the ELBO, the bias-variance decomposition, the dual form of the support vector machine, the DPO objective — it is given in full. Where a derivation does not matter, it is omitted and the result is stated with a pointer to where the derivation lives.
It is not a tutorial. There is very little runnable code, because runnable code goes stale faster than anything else in this subject and the official documentation will always be better at it. What you will find instead is the name of the function you want, the arguments that matter, and the trap you are about to walk into.