Lessons
Introduction to Machine Learning: all lessons
From what supervised learning is, to SHAP, K-means, recommenders and forecasting.
12 modules · 21 lessons · 1 free to read now
About this track: outcomes, tools and priceModule 1Getting Started
2 lessonsThree kinds of learning, and the two data distinctions that drive almost every later decision.
What separates the three is not the algorithm but the shape of the feedback available. One question, and its answer decides everything else.
Categorical vs Numerical, Structured vs Unstructured
14 minTwo distinctions that decide which algorithms are available and what preprocessing is mandatory.
Included with the track Simulation inside
Module 2Dataset Splitting, Cross-Validation, Data Leakage
2 lessonsBefore any algorithm. A bad split produces a beautiful number that lies.
Splitting Data, and K-Fold
18 minWhy one split is not enough, and why k = 5 is nearly always the answer.
Included with the track Simulation insideData Leakage
16 minThe number-one cause of a model that shines on a laptop and fails in production. Three forms, and how to spot the symptom.
Included with the track Simulation inside
Module 3Training, Evaluation, Bias–Variance, Tuning
3 lessonsHow to train, how to measure with the right metric, and the one trade-off that explains nearly every model failure.
Choosing the Right Metric
18 minAccuracy is the default metric and is almost always wrong. Here is how to choose properly.
Included with the track Simulation insideBias–Variance Trade-off
16 minOne framework that explains underfitting, overfitting, and why more data sometimes does not help.
Included with the track Simulation insideHyperparameter Tuning
14 minGrid search, random search, and why random search is nearly always better.
Included with the track Simulation inside
Module 4KNN — Classifier and Regressor
1 lessonThe algorithm that learns nothing, and teaches a great deal precisely because of that.
How KNN Decides
16 minNo training, no model. Only distance, and a k that decides everything.
Included with the track Simulation inside
Module 5SVM — Classifier and Regressor
1 lessonFind the boundary with the widest gap to both classes. Simple, and very strong in high dimensions.
Margin, C, and Kernels
20 minOne geometric idea, one parameter that sets tolerance, and one trick that lets it curve.
Included with the track Simulation inside
Module 6Linear Regression
1 lessonThe oldest model, the easiest to explain, and still a baseline that is hard to beat.
Finding the Best Line
18 min"Best" means the smallest sum of squared residuals. Why squares, and what follows from that.
Included with the track Simulation inside
Module 7Logistic Regression
1 lessonCalled regression, does classification. And it outputs a probability, which is more useful than a class.
Sigmoid, Log-Odds, and the Threshold
18 minWhy linear regression cannot do classification, and what the sigmoid changes.
Included with the track Simulation inside
Module 8Decision Tree, Random Forest, Gradient Boosting
1 lessonFrom one tree you can draw, to hundreds that win nearly every tabular competition.
From One Tree to a Forest
22 minOne tree is easy to read and easy to overfit. Two ways to combine many, and why they differ.
Included with the track Simulation inside
Module 9Feature Engineering
1 lessonThe work that most often decides the result, and is least often covered in tutorials.
Building Features That Help
20 minEncoding, transformation, date features, aggregation. Plus one warning about target encoding.
Included with the track Simulation inside
Module 10Explainable AI: Feature Importance, PDP, SHAP
1 lessonThree tools answering three different questions. Using the wrong one gives a confidently wrong answer.
Three Questions, Three Tools
22 minWhich features matter? What shape is the effect? Why was this applicant rejected? Three questions, and no single tool answers all three.
Included with the track Simulation inside
Module 11Unsupervised Learning: Clustering & Dimensionality Reduction
3 lessonsFinding groups without labels with K-means, and condensing many columns into a few with PCA.
Clustering with K-Means
20 minTwo steps on repeat — assign every point to its nearest centre, move each centre to the mean of its group — and you have the most widely used clustering algorithm.
Included with the track Simulation insidePrincipal Component Analysis (PCA)
20 minPCA finds the directions in which the data varies most and projects the data onto a few of them. Many correlated columns become a few uncorrelated components.
Included with the track Simulation insideDBSCAN, Hierarchical Clustering and Gaussian Mixtures
20 minClustering for groups that are not round, and when to use each method.
Included with the track Simulation inside
Module 12Recommendation Systems
4 lessonsFilling the blank cells of a user–item matrix three ways — demographic segments, item features, and similar neighbours — then measuring which one is right.
What a Recommender Predicts
18 minEvery recommendation is a guess at one blank cell in a user–item matrix — and the easiest guess, the segment average, turns out to be hard to beat.
Included with the track Simulation insideContent-Based Filtering
18 minIf you know what an item is made of, you can recommend it to people who liked similar items — even before anyone has rated it.
Included with the track Simulation insideCollaborative Filtering
22 minPeople who rated alike in the past will rate alike in the future — from that one assumption come user-based, item-based, and matrix factorisation.
Included with the track Simulation insideEvaluating and Combining Recommenders
20 minA smaller RMSE is not necessarily a better list: ranking metrics, A/B tests, hybrid strategies, and the popularity bias that quietly reinforces itself.
Included with the track Simulation inside