Reference¶

Material you come back to, rather than read once. Nothing here is assigned. Open a page when a week needs it, or when something in a lab stops making sense.

When you need what¶

Week	You are doing	Open this
04	Evaluation and error analysis	Choosing a metric
05	Linear models and the loss landscape	Linear algebra
06	Optimization and gradient descent	Calculus for ML
07	Bias, variance and regularization	Linear algebra, norms section
09	Model selection and calibration	Probability and statistics
11	Dimensionality reduction and PCA	Linear algebra, SVD section
any	Writing pipeline code	scikit-learn cheatsheet

If you are in Weeks 1 to 3, you can ignore this whole section. The maths arrives in Block 2.

Maths you actually use¶

Each page is scoped to the two or three weeks that need it, and says up front what it leaves out.

Linear algebra

Vectors, dot products, matrix shapes, norms, projection, and the SVD that PCA is built on.

Open
Calculus for ML

Derivatives, gradients, the chain rule, and deriving the update rule that gradient descent runs on.

Open
Probability & statistics

Distributions, expectation and variance, sampling, and deciding whether a difference between two models is real.

Open

While you work¶

Choosing a metric

Precision against recall, ROC against PR under imbalance, RMSE against MAE, and what each one hides from you.

Open
scikit-learn cheatsheet

Pipelines, column transformers, cross-validation, and the API patterns worth committing to memory.

Open

The data this course uses¶

Two datasets carry the whole course, and both are public. You never download either by hand: a script fetches them, which is the same habit the projects are graded on.

Dataset	Used in	Published by	Licence	How to get it
Olist Brazilian e-commerce, about 100,000 orders, 2016 to 2018	Block 1 labs, Weeks 1 to 4	Olist, on Kaggle	CC BY-NC-SA 4.0, so non-commercial use with attribution	`python data/download.py`
New York yellow taxi trips, a fixed 600,000-trip extract, May to July 2026	Project 1, and Block 2 from Week 5	NYC Taxi and Limousine Commission, as open data	Published by the City of New York under its open data terms	`python data/download.py --taxi`

Both commands read from the course data releases, where the exact files are pinned so that every student works with identical rows. The taxi release also carries the official data dictionary and the zone lookup table.

Why this table exists

Your own PROBLEM.md is asked to record where your data came from, who published it, and under what terms. It seemed fair to do the same here.

Where to find data¶

For the capstone, and for anything you build after this course. Your project dataset has to meet the requirements in the project brief, so read those before you fall in love with something.

Kaggle Datasets. Enormous and uneven. Filter hard, and avoid anything with a thousand published notebooks.

UCI Machine Learning Repository. Small, clean, well documented. Good for drills, usually too tidy for a project.

EU Open Data Portal and INSTAT. Real administrative data, genuinely messy, and local.

Hugging Face Datasets. Strong for text and images, with licences stated properly.

Google Dataset Search. Use when you know the domain but not the source.

Going further¶

Five things worth your time, in rough order of how soon they are useful.

3Blue1Brown, Essence of Linear Algebra. Fifteen short videos. The best intuition for matrices anyone has built. Watch it even if you did linear algebra already.

Mathematics for Machine Learning, Deisenroth, Faisal and Ong. Free PDF. The maths pages here are a scoped version of its first five chapters.

Hands-On Machine Learning, Géron. The closest thing to a textbook for this course. Chapters 1 to 7 track Blocks 1 and 2 closely.

An Introduction to Statistical Learning, James et al. Free PDF. Stronger than Géron on why methods work, lighter on code.

scikit-learn user guide. Not a fallback for when documentation fails. It is genuinely one of the better explanations of applied machine learning in existence.