Machine Learning

How models actually learn, explained with intuition first and a little NumPy to prove it. 18 modules of 10 minutes each, with questions that check your answers in the browser.

Start the track

Foundations

• 1. What ML is: supervised vs unsupervisedFree
  See machine learning as "finding a rule from examples", tell supervised from unsupervised problems, and know when not to use ML at all.
• 2. Features, labels, training vs test dataFree
  Organise data into a feature matrix X and a label vector y, encode categories, split into training and test sets, and avoid leaking the answer into the features.
• 3. The idea of a model and a loss functionFree
  See a model as a function with parameters, measure how wrong it is with a loss function, and find the best parameters by searching for the lowest loss.
• 4. Overfitting, underfitting, and generalizationFree
  See how a model can be too simple or too flexible, measure it with training vs test error, and learn the habits that keep models honest.

Intermediate

• 5. Linear regression (the math intuition)Free
  Fit a line (and a plane) that minimises squared error, read the coefficients as business effects, compare NumPy's least squares with scikit-learn, and check the residuals.
• 6. Logistic regression and classificationFree
  Turn a weighted sum into a probability with the sigmoid, predict churn with logistic regression, choose a decision threshold, and read the coefficients.
• 7. Train/validation/test split and cross-validationFree
  Use a validation set to choose models without touching the test set, get more reliable estimates with k-fold cross-validation, and tune settings with a grid search.
• 8. Classification metrics: accuracy, precision, recall, F1Free
  Count true and false positives, compute precision, recall and F1 by hand and with scikit-learn, and choose the metric that matches the business cost.
• 9. Regression metrics: MAE, MSE, R²Pro
  Measure regression errors in units people understand, know when MSE or MAE fits, interpret R² correctly, and always compare with a naive baseline.
• 10. The bias-variance tradeoffPro
  Understand prediction error as bias plus variance plus noise, see both by refitting models on many samples, and use learning curves to decide what to fix.

Advanced

• 11. Gradient descent explainedPro
  Train a model by repeatedly stepping downhill on the loss, compute gradients for linear regression by hand, choose a learning rate, and see why it powers neural networks.
• 12. Regularization (L1, L2)Pro
  Fight overfitting by penalising large weights, compare Ridge (L2) and Lasso (L1), tune the penalty strength with cross-validation, and use Lasso to select features.
• 13. Decision trees and random forestsPro
  Understand how a tree splits data with yes/no questions, read a fitted tree, control its depth, and combine many trees into a more accurate random forest.
• 14. Feature scaling and preprocessingPro
  Standardise numbers, one-hot encode categories, fill missing values, and wrap it all in a Pipeline with ColumnTransformer so training and prediction always match.
• 15. Confusion matrix and ROC/AUCPro
  Lay out every type of mistake in a confusion matrix, trace precision-recall and ROC curves across thresholds, and summarise ranking quality with AUC.

Expert

• 16. Ensemble methods and boosting (intuition)Pro
  Combine models by voting, bagging and boosting, understand how gradient boosting corrects its own mistakes step by step, and know when it's worth the extra tuning.
• 17. Unsupervised: clustering and PCAPro
  Find customer segments with k-means, choose the number of clusters, describe what each segment means, and compress many features into a few with PCA.
• 18. A clean ML workflow end to endPro
  Go from business question to a deployed, documented churn model with a reproducible pipeline, honest evaluation, a chosen threshold, a saved model and a scoring function.

Track project

• Machine Learning Project: Toko Kita's first modelsPro
  Three graded tasks on the shop's real data. Build a churn model and evaluate it honestly, choose the decision threshold that costs the shop least, and pick a demand forecaster with cross-validation instead of a lucky split.

Pro modules unlock with one payment of Rp 149.000, yours to keep.