Advanced modules

Nine optional, self-study modules. Four take the course beyond ISLP: causal inference with randomised experiments, explaining black-box models with Shapley values, distribution-free uncertainty with conformal prediction, and the exponential-family machinery that unifies GLMs and penalized splines. Five more are ISLP chapters lifted out of the taught sequence — support vector machines (Ch 9), deep learning (Ch 10), survival analysis (Ch 11), unsupervised learning (Ch 12) and multiple testing (Ch 13).

9 modules 688 core slides (+89 optional) 54 + 23 exercises with solutions 9 companion notebooks optional self-study

Each module is built exactly like a course deck — the same colour-coded callout boxes, every exercise followed immediately by its worked solution, a closing summary block and an optional appendix — and each is paired with a companion notebook whose printed numbers match the slides seed-for-seed (np.random.default_rng(2024)). The modules sit outside the taught plan: nothing in the course depends on them. Note that the mock exams still carry a multiple-testing problem, which now draws on module A7, and any unsupervised-learning question now draws on module A8.

The modules

	Module	What it covers	Slides	Deck	Notebook	Colab
A1	Randomised Controlled Trials	Potential outcomes, the selection-bias decomposition, why randomisation works, regression adjustment with HC2 errors, power and sample size, ITT vs per-protocol, the peeking problem	69 (+6)	Open	Rendered
A2	Explainable AI with Shapley Values	Cooperative games and the four axioms, the marginal value function, exact Shapley by enumeration, Monte-Carlo sampling, waterfalls and global importance, the correlated-feature and retrain pitfalls	71 (+8)	Open	Rendered
A3	Conformal Prediction	Exchangeability, split conformal and the finite-sample quantile, marginal vs conditional coverage, locally weighted scores and CQR, classification prediction sets, the OLS stress test	70 (+7)	Open	Rendered
A4	GLMs and Splines	The exponential family, Poisson regression on Bikeshare, deviance and LRTs, overdispersion (quasi-Poisson, negative binomial), penalized splines and effective df, a count GAM	82 (+8)	Open	Rendered
A5	Support Vector Machines	The maximal-margin classifier, the soft margin and the cost C, the support-vector classifier, polynomial and radial kernels, tuning by cross-validation	84 (+10)	Open	Rendered
A6	Survival Analysis	Censoring and why it breaks ordinary regression, the survival and hazard functions, Kaplan–Meier, the log-rank test, Cox proportional hazards	86 (+10)	Open	Rendered
A7	Multiple Testing	Why naive testing fails at scale, FWER, Bonferroni and Holm, the false discovery rate, Benjamini–Hochberg, p-hacking	63 (+11)	Open	Rendered
A8	Unsupervised Learning	No test error to validate against, principal components (loadings, scores, PVE, the biplot), the scaling decision, K-means and its local optima, dendrograms, linkage and dissimilarity, clusters in pure noise	91 (+10)	Open	Rendered
A9	Deep Learning	Single-layer networks and activations, MLPs and parameter counts, convolutions and pooling, loss and SGD, regularisation and dropout	72 (+19)	Open	Rendered

Tip

Everything is implemented from scratch — exact Shapley values, split conformal, P-splines — so the notebooks need nothing beyond the course environment. The slides name the production tools (shap, MAPIE, mgcv-style GAM software) once the mechanics are understood.

What each module assumes

Each module bridges back to the course chapters it extends, and assumes them:

Module	Builds on
A1	Ch 0 (standard errors, tests) · Ch 3 (regression) · Ch 5 (simulation) · A7 (multiple testing — the peeking problem)
A2	Ch 0b (counting and the 2ᵖ cost) · Ch 8 (boosting, variable importance) · A9 (black-box models)
A3	Ch 3 (prediction intervals) · Ch 5 (train/validation splits)
A4	Ch 3–4 (linear and logistic regression) · Ch 5–6 (CV, AIC) · Ch 7 (splines and GAMs)
A5	Ch 4 (classification, ROC/AUC) · Ch 5 (cross-validation)
A6	Ch 0 (distributions, tests) · Ch 3 (regression) · Ch 4 (logistic regression)
A7	Ch 0 (hypothesis testing, p-values) · Ch 5 (resampling)
A8	Ch 2 (what a test error buys you) · Ch 5 (resampling and stability)

How they are built

The sources live in Chapters/Advanced/ — one folder per deck with its LaTeX source, its make_figures.py (every figure is computed from the course datasets or a seeded simulation) and the compiled PDF; the notebooks ship with stored outputs and run on Colab with one click, resolving data exactly like the course labs. From the repository root, make advanced rebuilds any deck whose source changed. The folder also carries the distilled style guides (STYLE_DECK.md, STYLE_NOTEBOOK.md) used to author them — the starting point for adding a module A9.

All advanced notebooks

• Advanced Module A1 — Randomised Controlled Trials
• Advanced Module A2 — Explainable AI with Shapley Values
• Advanced Module A3 — Conformal Prediction
• Advanced Module A4 — GLMs and Splines
• Chapter 9 — Support Vector Machines
• Chapter 11 — Survival Analysis
• Chapter 13 — Multiple Testing
• Chapter 12 — Unsupervised Learning
• Advanced Module A9 — Deep Learning

Where to go next

• Lecture slides — the ten course decks these modules extend.

• Lab notebooks — the course labs, built to the same pattern.

• The course at a glance — where the taught plan ends and these begin.

• Repository layout — where everything lives.