Learn Classification & Regression with Phoebe

The supervised zoo: seven classic algorithms, in depth.

Intro-ML taught you the workflow with one classifier and one tree. This 10-session builder course opens the whole cage: linear and polynomial regression, logistic regression from the inside (odds, coefficients, log-loss), gradient descent built from scratch, k-nearest neighbors, naive Bayes, decision trees in real depth (gini, entropy, pruning), and support vector machines - then two sessions on the judgement call that matters most: choosing the right algorithm, and proving your choice in a full bake-off. Same Lumen Skincare data, same seed 42, every number validated with real scikit-learn runs - so what you learn here reconciles line by line with the statistics, intro-ML, and experimentation courses.

10sessions, one track
7algorithms, in depth
4boundary styles, live
2leaderboards to win

The sessions 🛠️ · for practitioners · 10 x 45 min · Python + scikit-learn · assumes intro-ml

One algorithm per session, one running brand across all ten. Every session ends the same way: the algorithm faces the Lumen data, earns its spot on the leaderboard, and you learn exactly why it landed where it did. Everyone starts at session 1.

🗺️ Session 1 · start here

The supervised zoo

Seven algorithms, two families - model-based vs instance-based - and the one picture that compares them all: the decision boundary, explored live. Plus the baseline leaderboard the course will explain.

▶ Start hereLive boundaries 📏 Session 2

Linear regression, properly

Least squares from residuals up, coefficients you can read in dollars, R² honestly - then polynomial features and the degree-12 explosion that teaches restraint. Ridge and lasso as the tamer.

Least squares 🎚️ Session 3

Logistic regression, from the inside

Sigmoid, odds and log-odds, coefficients as odds ratios (email x4.1, returning x2.8), why the loss is log-loss and not MSE - and the moment the model recovers the machine that made the data.

Odds & coefficients ⛰️ Session 4

Gradient descent

The loss surface, the learning rate, batch vs stochastic - then the payoff: a 15-line from-scratch descent that matches scikit-learn's coefficients to two decimals. You will never fear "solver" again.

From scratch 📍 Session 5

k-nearest neighbors

The algorithm with no training step: k from memorization (k=1, a coin flip on test) to over-smoothing, why scaling is life or death (0.64 vs 0.58 AUC), the curse of dimensionality, and kNN regression.

Distance & k 🎲 Session 6

Naive Bayes

Bayes' theorem as a classifier, the independence assumption that is false yet works, Gaussian vs Multinomial vs Bernoulli, smoothing - and why you never trust its probabilities, only its rankings.

Probability engine 🌳 Session 7

Decision trees, in depth

Gini vs entropy (near-identical - gini is just cheaper), information gain, the depth U-curve, pruning with max_depth / min_samples_leaf / ccp_alpha, feature importance - and where trees lose to lines.

Splits & pruning 🥅 Session 8

Support vector machines

The widest-street idea, support vectors, C and gamma, the kernel trick - and the honest Lumen result: linear SVM ties logistic while a default RBF loses badly. Kernels are not magic.

Margins & kernels 🧭 Session 9

Choosing your model

Bias and variance across the whole zoo, the decision guide (data shape, size, interpretability, speed), and the fair-comparison protocol: tune by cross-validation, spend the test set once.

The judgement 🏆 Session 10

Capstone: the bake-off

Every algorithm, both Lumen tasks, one honest leaderboard each - written up as model cards and presented like an analyst (dollars before scores). Plus the bridges to what comes next.

Capstone
start here - the map & lines building up - the core zoo getting real - trees, SVM, judgement hardest - the bake-off

Choose your path 🗺️

Everyone starts at session 1. The fast-tracks fit a specific gap - but the full journey is the point.

🛠️ Full journey (recommended) 1→ 2→ 3→ 4→ 5→ 6→ 7→ 8→ 9→ 10
📈 Regression fast-track 1→ 2→ 4→ 7
🎯 Classification fast-track 1→ 3→ 5→ 6→ 7
🧭 "Which model?" crash course 1→ 9→ 10

Try the boundary explorer 🧪

Real models, fit live in your browser - real gradient descent, real CART splits, real neighbor votes. Switch algorithms and datasets, drag the knobs, and watch the train-test gap tell you when a model is memorizing instead of learning. The same tool is embedded in the sessions that teach each algorithm.

One brand, every algorithm. Every session works on Lumen Skincare - the synthetic $18M direct-to-consumer skincare label shared with the statistics, intro-ML, and experimentation courses (same generator, seed 42). Two tasks all course long: a classifier for whether a checkout session converts (~3% base rate) and a regressor for order value (mean ~$73, right-skewed). Every teaching number on every page was validated with a real scikit-learn run - and the honest headline is that on this tabular, mostly-linear data, the linear models top both leaderboards. Knowing why is the course.
Honest about scope. Sessions distill the classic supervised-learning canon - StatQuest's algorithm series, the scikit-learn user guide (linear models, SVM, nearest neighbors, naive Bayes, trees), and Andrew Ng / DeepLearning.AI's ML Specialization C1 + the C2 tree module. Certificates, videos, and graded labs stay official. The workflow itself (splits, CV, metrics) is assumed from learn-intro-ml-with-phoebe; feature engineering, ensembles/XGBoost, unsupervised learning, and deep learning are pointed to their own sibling courses. Content verified July 2026.

The knowledge map 🧠

The whole course at a glance - hover a session to spotlight its concepts, click any node to jump in.