Lesson sources and adaptation notes

Reviewed 14 October 2026. Teaching narratives, depot scenarios, hand-worked examples, exercises, decision costs and assessment design are original instructional material. Sources below support technical definitions, API behavior and statistical cautions; they are not evidence of real depot performance or measured learner outcomes. Numerical reference results come from the supplied runnable code and output files.

Version discipline

The runnable environment pins scikit-learn 1.8.0 and pandas 2.2.3. Versioned documentation is authoritative for implementation details. On the review date, moving stable documentation resolved to newer scikit-learn and pandas releases; those pages were consulted only for broad concepts also checked against the pinned APIs. No claim is made that the course uses the newest released package. Verify installed versions with the setup workflow.

Lessons 01–02: question, targets and provenance

• scikit-learn 1.8 DummyRegressor: constant mean/median strategies and the role of a simple reference estimator.
• scikit-learn common pitfalls, with 1.8 counterpart: information leakage, training-only fitting and consistent transformations.
• pandas 2.2 DataFrame.merge: relationship validation, match indicators and the warning that null keys may match each other.
• pandas merging user guide: broad join/concatenation context; exact course behavior is tied to the versioned API above.

The prediction-time contract, maturity ledger and example completion/confirmation timelines are original applications of information-availability reasoning. The synthetic data assume completion and label availability coincide in label_available_at; a real source's measurement and reporting behavior would need independent verification.

Lessons 03–04: validation design and preparation

• scikit-learn 1.8 TimeSeriesSplit: expanding training portions, default validation sizes, sample-based gaps and equal-spacing caution.
• Cross-validation guide, with 1.8 counterpart: random, time-aware and grouped designs address different dependence and generalization questions.
• ColumnTransformer: explicit column routing, combination of transformed outputs and remainder behavior.
• OneHotEncoder: categorical representations, fitted categories and unknown handling.
• StandardScaler: centering/scaling learned from training data and population-standard-deviation convention.
• SimpleImputer: fitted replacement statistics and missingness-indicator behavior.

The exact day allocation is a course protocol, not a default recommended by a library. Six arrivals occur each day, but overnight and within-day spacings differ. A 12-row gap therefore corresponds to two days' worth of omitted jobs; timestamp checks, not row counts alone, establish maturity. Inner validation boundaries can fall within a day.

Lessons 05–06: supervised families and evaluation

• Linear models and Ridge: regularized linear prediction and the role of the penalty.
• Forest ensembles: aggregation of randomized trees and their model-family characteristics.
• LogisticRegression: classifier API, regularization, class weights and probability output.
• Model-evaluation guide: regression errors, confusion-based measures, average precision, score direction and evaluation caveats.
• Log loss: probability-loss definition and label conventions.
• ROC AUC: ranking-based score and input requirements.
• NIST/SEMATECH: checking model fit: residual inspection as an essential complement to numerical fit summaries.
• An Introduction to Statistical Learning, authors' site: supplementary open statistical-learning reading, especially the concepts of supervised learning, validation and model criticism. The course does not reproduce textbook passages or exercises.

Metrics are calculated explicitly in original miniatures. No hand-worked value is presented as an observed metric from the generated dataset. Predictive association, feature importance and residual patterns are not interpreted as causal intervention effects.

Additional versioned sources for Lessons 07–10

Reviewed on 14 October 2026. These are official scikit-learn 1.8.0 pages, matched to the course's pinned scikit-learn version rather than a moving latest-version page. They support API behavior and technical cautions, not measured model results. Lesson examples, synthetic costs, teaching exercises and the depot split protocol are original instructional constructions. Actual dataset metrics must come from outputs/reference_results.json.

Lesson 07

• TimeSeriesSplit: expanding ordered training sets; default fold size; gap measured in samples; equal-spacing caution. Fold indices in the lesson are calculated from the API contract.
• Cross-validation guide: selecting validation strategies for time-dependent and grouped observations; separating model selection from evaluation.
• Common pitfalls: preprocessing leakage, pipelines inside cross-validation, and random-state behavior.
• GridSearchCV: parameter-grid fitting, scoring and refitting. Negative losses follow scikit-learn's greater-is-better scoring convention.

Lesson 08

• Probability calibration: reliability diagrams, probability-loss interpretation, calibration methods, disjoint data for calibrating a fitted FrozenEstimator and small-sample isotonic caution.
• Binary Brier loss: average squared probability error for the binary 0/1 outcome used in the hand traces. The lesson uses the binary 0–1 scaling.
• Decision-threshold tuning: distinction between probabilities and decisions, decision metrics and avoiding threshold tuning on model-fitting observations. The specific depot costs and capacity cap are course assumptions, not external operational evidence.

Lesson 09

• Mean absolute error: loss definition and units for regression examples.
• Permutation feature importance: score deterioration interpretation, dependence on fitted model and evaluation data, and misleading low importance with correlated inputs.
• Permutation-importance API: score choice, repeated permutations and fitted-estimator behavior. Repetition of shuffles is not presented as independent outcome sampling or causal inference.

Lesson 10

• Common pitfalls: randomness: reproducibility implications of random-state choices. The course uses documented fixed settings for a repeatable instructional run; repeatability is not a robustness claim.
• Model persistence: arbitrary-code risks of pickle-based formats and the need to preserve a compatible software environment.

The drift arithmetic and handoff checklist are original extensions of the course's data-quality, denominator and reproducibility contracts. They describe a bounded monitoring design exercise, not a deployed monitoring system or a claim of statistically validated drift detection.

Prerequisite alignment inspected locally

The author reviewed the sibling Python, SQL and Statistics with Data Storytelling syllabi. Lessons build on Python functions, NumPy/pandas, tests and reproducible outputs; SQL grain, cutoff and leakage-aware feature extraction; and statistics residuals, denominators, uncertainty and causal restraint. No clustering, deep learning or high-impact scoring is introduced.

Lessons 11–12: integration and assessment

The integration assembles the preceding sources into an original synthetic workflow. The calibration and permutation-importance documentation support interpretation cautions; source and data hashes, freeze records and learner-owned evidence provide an instructional reproducibility protocol. These checks are transparent and do not create a security boundary around physically accessible test files.

The 100-point rubric, integrity gates, 80-point completion suggestion, 36 exercises, 24 exit checks and 30–42-hour planning range are proposed instructional design. No source validates those thresholds or timings; no learner pilot or accreditation is claimed. The formative 72-point self-check scale is separate from the capstone rubric.

Prerequisite continuity

The following local Aabha syllabuses were inspected during authoring:

• Python for AI and Data: functions, modules, CSV validation, debugging/tests, NumPy, pandas and honest charts
• SQL for Data Analysis and AI: row grain, joins, duplicates, cutoff-aware feature extraction and leakage
• Statistics with Data Storytelling: denominators, sampling/dependence, uncertainty, residuals, effect interpretation and causal restraint

This course adds learned preprocessing, temporal model selection, predictive evaluation, probability calibration, threshold policy and reproducible handoff. It deliberately excludes clustering, neural networks and high-impact scoring. Deep learning belongs to the separate follow-on course rather than being introduced as an obligatory improvement over checked tabular models.

Source access snapshot

Checked on 2026-10-14. The moving stable scikit-learn pages returned version 1.9.1 and pandas stable documentation returned 3.0.6 during authoring. The executed course deliberately uses verified older pins scikit-learn 1.8.0 and pandas 2.2.3, with matching versioned API links. Package availability was verified through installation and execution, not inferred from page titles. External references are linked, not redistributed.