Recap and final exam
Ten modules, one conviction: performance is won in the features. Here is the course condensed — module by module first, then through the threads running across it.
The course at a glance
| Module | The essential takeaway |
|---|---|
| 1. Features or algorithm | A feature is a representation; a model only learns what the space makes expressible |
| 2. Missing values | Ask why it is missing (MCAR, MAR, MNAR) before how to fill; the indicator keeps the signal |
| 3. Scaling | Needed for distances, dot products, gradient; useless for trees; does not fix shape |
| 4. Categorical encoding | Numbering creates a false order; one-hot at low cardinality, target with smoothing beyond |
| 5. Dates and cycles | Decompose a date into ten features; sine-cosine to bring December and January together |
| 6. Text features | TF-IDF by default, n-grams for order; do not neglect length and punctuation |
| 7. Aggregations and windows | Deviation from habit beats the raw value; shift(1) before rolling |
| 8. Data leakage | Improves the score, hence invisible to metrics; four families, one question to ask |
| 9. Selection | Filters to trim, permutation to decide; operating cost counts |
| 10. Pipelines | A Pipeline makes leakage structurally impossible and removes the production gap |
The threads running through the whole course
Fit on training, apply everywhere. The same sentence recurs in modules 2, 3, 4, 6 and 10. Every learned statistic — median, mean, standard deviation, per-category means, vocabulary, IDF weights — is computed on training data alone. This is not one precaution among others: it is the boundary between a real score and an invented one.
The availability question, asked of every feature. "Is this information already known at the moment I must predict?" It comes from the supervised course, structures module 8, and disqualifies the temporal features of module 5 as much as the aggregations of module 7. A highly predictive feature that is unavailable in production is not an asset: it is a leak.
Relate a value to its relevant reference. This is the mechanism that produces the most gains: an amount compared to the customer's habit (module 7), a category replaced by its conversion rate (module 4), a value standardized per group rather than over the population. The raw value rarely says as much as the deviation.
The choice of transformation depends on the intended model. Scaling is useless for trees but indispensable for PCA and regularized models; cyclical encoding is precious for a linear model, marginal for a tree; high cardinality is handled natively by LightGBM. There is no universal preprocessing, only preprocessing suited to a purpose.
The final exam
The exam comprises 40 questions covering the ten modules: missingness mechanisms and imputation strategies, choice of scaling and distribution transformations, encoding by cardinality, temporal and cyclical features, text representations, aggregations and windows, identifying leaks, selection methods and building pipelines.
Several questions present mini-scenarios: a suspicious feature you must judge for leakage, a dataset whose encoding you must choose, an abnormally high score whose origin you must find. The goal of this course is judgment, not recitation.
On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.
Reread the table above, then for each line ask: "and if I got this wrong, how would I notice?". If the answer comes — why mean imputation biases under MNAR, why a rolling without shift destroys a temporal model, why an excellent score should worry you — 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 examYou need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.