Recap and final exam

Ten modules, one complete method. Before the exam, here is the course condensed — module by module first, then through the cross-cutting threads that connect them.

The course at a glance

Module	Core takeaway
1. Framing	Target decides regression vs classification; features must exist at prediction time; sealed test set
2. Linear regression	Weighted sum, readable coefficients; Ridge shrinks, Lasso selects
3. Logistic regression	Sigmoid turns score into probability; the threshold trades precision against recall
4. k-NN and SVM	Neighbor vote (standardize!) vs maximum margin; RBF kernel for non-linear boundaries
5. Decision trees	Readable yes/no questions; always overfit alone; contained by depth and leaf constraints
6. Random forests	Bagging + random features = decorrelated trees; variance divided; free out-of-bag score
7. Gradient boosting	Sequential trees correcting residuals; low learning rate + early stopping
8. Validation	Stratified k-fold; temporal splits for time data; pipelines make leakage impossible
9. Metrics	Accuracy lies when imbalanced; precision/recall per business cost; ROC AUC across thresholds
10. Tuning	Random search beats grid at equal budget; test set touched exactly once

The threads that run through everything

Bias and variance, everywhere. Every dial in this course moves the same slider: $\lambda$ in Ridge/Lasso, $k$ in k-NN, C in SVM, max_depth in trees, learning_rate in boosting. Recognizing which side of the trade-off you're on — underfitting or overfitting — tells you which way to turn any of them.

Two opposite ensemble philosophies. The forest trains strong trees in parallel and averages away their variance; boosting trains weak trees in sequence and grinds down the bias. Same ingredient, opposite assembly — and together they dominate tabular machine learning.

Honest measurement is the real skill. Models are interchangeable; method is not. Sealed test set, stratified or temporal cross-validation, pipelines against leakage, metric chosen from business cost before training: this discipline is what separates a real result from a beautiful illusion.

Simplicity is a strategy. The baseline of module 10 is not a warm-up: a regularized linear model sets the bar, sometimes holds it, and always tells you what the added complexity is worth.

The final exam

The exam includes 40 questions covering the ten modules: problem framing, linear and logistic models, geometric methods, trees and ensembles, validation, metrics and tuning. Several questions present mini-scenarios — an imbalanced dataset, a suspicious score, a temporal split to design — because the course's goal is judgment, not recitation.

On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.

Before you start

Reread the summary table above, then ask yourself for each line: "why?". If the answer comes — why does accuracy lie when classes are imbalanced, why does Lasso zero coefficients, why must preprocessing live inside the pipeline — you are ready. Good luck!

Final exam

Ready to validate this course?

40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.

Start the exam

You need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.