ModernCS
Join for free
scikit-learn

Intelligence · Taught with scikit-learn

Machine Learning

Regression, trees and gradient boosting on real tabular data. You build models that predict honestly, measure them properly, and learn to spot the leakage that makes bad results look good.

Download the course plan
  • Foundational
  • 4 phases
  • 21 sessions
  • First session free

Syllabus

4 phases · 21 sessions · each phase ends in something you have built

1

Before You Fit Anything

Build the numeracy this course actually uses, and beat guessing without a model.

$72

this phase

  1. 1.1

    What a Prediction Problem Is

    Free preview

    Rows, features, a target, and framing a question someone actually has.

    Open
  2. 1.2

    Tables in Python

    Load a CSV, select rows and columns, check types, and find what is missing.

  3. 1.3

    Just Enough Statistics

    Mean, spread, base rate, and probability as a number you can already predict with.

  4. 1.4

    Just Enough Linear Algebra

    A row is a vector, a weighted sum is a dot product, and that is most of a model.

  5. 1.5

    Predicting Without Machine Learning

    Write a rule by hand, measure its error, and set the number every model must beat.

By the end of this phase:A data profile and a hand-written rule with a measured error, written before any model exists.
2

The Supervised Loop

Train a model, measure it honestly, and know when the measurement is a lie.

$72

this phase

  1. 2.1

    Fit, Predict, Score

    The three calls every scikit-learn estimator shares, run on a real dataset.

  2. 2.2

    Held-Out Data and the Dummy

    Train/test splits, DummyClassifier, and why 92% accuracy is often worthless.

  3. 2.3

    Linear and Logistic Regression

    A weighted sum, a loss, downhill, and reading what the coefficients claim.

  4. 2.4

    Overfitting and Underfitting

    Capacity, learning curves, regularization, and recognizing memorization on sight.

  5. 2.5

    Cross-Validation and the Right Metric

    Confusion matrices, precision, recall, ROC, and choosing the metric before you train.

By the end of this phase:A cross-validated model with a written defense of its split and its metric.
3

The Toolkit That Wins

Learn the models that actually win on tables, and compare them without cheating.

$72

this phase

  1. 3.1

    Decision Trees

    Splits you can read out loud, and watching a tree memorize its training set.

  2. 3.2

    Forests and Gradient Boosting

    Bagging, boosting, HistGradientBoosting, LightGBM, and why they still own large tables.

  3. 3.3

    Pipelines and Leakage

    ColumnTransformer, encoding, scaling, and fitting on training data only.

  4. 3.4

    Feature Engineering

    Dates, counts, ratios, and TF-IDF text features that beat a fancier model.

  5. 3.5

    Tuning Without Fooling Yourself

    Randomized search, nested validation, and spending your sealed test set once.

  6. 3.6

    Tabular Foundation Models

    TabPFN-3 against a tuned booster on your data, plus its size, cost, and license limits.

By the end of this phase:A leakage-free pipeline comparing four model families under one harness on a sealed test set.
4

Messy Data, Real Users

Handle the data you actually get, then put a model in front of people and defend it.

$72

this phase

  1. 4.1

    When There Is No Label

    k-means, PCA, anomaly detection, and what a cluster does not prove.

  2. 4.2

    Imbalance, Missingness and Time

    Rare classes, missing values, and splits that respect time and groups.

  3. 4.3

    Explaining One Prediction

    Permutation importance, partial dependence, SHAP, and an answer a non-engineer accepts.

  4. 4.4

    Serving Without Skew

    An API that loads the exact fitted pipeline you validated, not a retyped copy.

  5. 4.5

    Drift, Fairness and the Refusal

    Retraining triggers, error broken out by subgroup, and when the answer is do not ship.

By the end of this phase:A deployed model API with a model card, a subgroup error table, and a monitoring plan.

Tools you will use

  • Python
  • NumPy
  • pandas
  • scikit-learn
  • LightGBM
  • TabPFN-3
  • SHAP
  • FastAPI

What you will build

  • The Rule You Have to Beat

    A hand-written baseline, measured before any model

  • Model Bake-Off

    Logistic, forest, booster, TabPFN-3, one harness

  • Deployed Risk Scorer

    Served pipeline with a model card and drift alerts

Machine Learning · ModernCS