1. Open a free notebook at colab.research.google.com (sign in with a Google account and click New notebook). Think of it as a page in your browser where little blocks of code run when you press play—you never write code from scratch, you just copy, paste, and run.
2. Come back here, click “Start here” below, and follow the foundation page once, top to bottom. It sets everything up and comes ready to run with a built-in example, so you can watch it work before changing a thing.
3. After that, pick any model lesson that interests you and add its blocks to the same notebook. There's no fixed order—go at your own pace, and there's nothing you can break.
Once the foundation is done, pick any lesson that interests you
Your problem, your data
Start here. Decide what one row stands for, pick the answer you want to predict and the clues you'll use, load the example (or your own file), and set the baseline that every model will be measured against.
K-Nearest Neighbours
The "ask a few similar people" model: it predicts by looking at the rows most like yours. You'll see why putting all the clues on a fair scale matters, and how to choose how many neighbours to ask.
Linear Regression
Predict a number by giving each clue its own weight and adding them up—simple enough to read like a recipe. You'll learn to see which clues matter most, and how to avoid accidentally letting the answer sneak into the clues.
Logistic Regression
Sort each row into a category (yes/no, pass/fail) and get a sense of how confident the model is. You'll learn to read those confidence levels—and why "80% likely" is a leaning, not a promise.
Decision Trees
Like a game of twenty questions: the model learns a chain of simple yes/no questions to reach an answer. You'll see how to stop it before it starts memorising instead of truly learning.
Random Forests
Ask a whole crowd of decision trees and let them vote—a crowd is usually wiser than any one member. You'll also learn which clues the crowd leaned on, and why "leaned on" doesn't mean "caused".
Gradient Boosting
Build small trees one after another, where each one fixes the mistakes of the last—learning by patient correction. Then compare every model you've tried, side by side, in your own results log.
K-Means
No answer column to predict? This finds natural groups hiding in your rows—and shows you how to tell whether the groups are meaningful when there's no answer key to check against.
Do the foundation once, then add each new lesson underneath it in the same notebook. If Colab ever forgets where it was, just click Runtime → Run all and it picks right back up—nothing is lost.
Your numbers won't (and shouldn't) match anyone else's—they depend on your data. There's no single right answer here. The only fair comparison is against your own baseline, and how much a mistake would cost you.
You tune a model using cross-validation on the practice data. The set-aside test data is like the final exam—you look at it once, at the very end, so the score stays honest.