Curriculum Outline

Course week (Days 1–4)

Teaching decks on this site: Day 1 (Monday), Day 2 (Tuesday), and Day 4 (Thursday). Day 3 (Wednesday) is a mini symposium and social event — no lecture deck here.

This outline links to afternoon lab exercises (numbered tasks and one solution notebook per teaching day) while weaving a few ideas through the week: two parallel examples (simulated genes where we know the truth + Palmer Penguins as one real table), what a wrong prediction costs (thresholds, prevalence), calibration, “correlation is not causation” when reading importance plots, short data and model summaries, reproducibility, fairness-style subgroup checks, and assignments that reward honest wording (especially around SHAP).

Integrated design principles (built into the week)

  • Two parallel data tracks — Synthetic genes in the original linear / logistic notebooks: we scripted the biology so we know which coefficients should matter. Palmer Penguins (data card, catalog, dedicated notebooks, Tuesday’s full modeling walk-through in R using tidymodels; reading companion for Chapters 4, 7, and 8): one small real table reused Mon→Thu for plots, regression, classification, and workflows.
  • Decision costs before a long list of metrics — introduce thresholds, prevalence, and misclassification costs early; return to them on Thursday so ROC / PR / F1 choices feel motivated, not arbitrary.
  • Association ≠ intervention — embed a short “causal guardrails” beat when introducing importance and SHAP so plots are not read as policy levers by default.
  • Data cards + model cards — start a one-page data card Monday (where the numbers came from, leakage risks, subgroup fields); finish a model card Thursday (intended use, metrics, limits, monitoring).
  • Reproducibility spine — Tuesday includes a concrete folder layout + “one command rebuilds figures” goal (quarto project; optional renv / conda lockfile for a reproducibility homework sprint).
  • Fairness as a metric choice — Thursday explicitly compares overall vs subgroup metrics on Palmer Penguins (e.g. species or island as a subgroup lens) when a sensitive attribute exists (or simulate site/batch as a stand-in).
  • Assessment that teaches humility — include a short oral or written prompt: “What would you not claim from this SHAP summary?” graded with a published mini-rubric (correctness, diagnostics, communication).

Day 1 (Monday): Foundations and Regularization

Learning goals

  • Understand supervised learning setup for regression and classification on genes (simulation) and Palmer Penguins (real)
  • Use linear and logistic regression to frame overfitting vs underfitting
  • Learn why regularization is needed and when to use ridge vs lasso
  • Draft a data card (sources, leakage, subgroups) and connect it to train/validation design

Sequence

  1. Framing predictive modeling: data split, validation, and generalization (tie each choice to the data card)
  2. Linear regression as baseline and bias–variance discussion (synthetic gene–trait story in notebooks; penguins for real EDA in parallel)
  3. Logistic regression for classification intuition (same patient/scientific framing as Thursday metrics)
  4. Overfitting/underfitting diagnostics (learning curves; train vs validation) on a held-out split (genes or penguins)
  5. Regularization: ridge and lasso; cross-validation (LOO and (k)-fold figures) to choose (); shrinkage paths; preview elastic net — Module 02 + Day 1 CV slides
  6. Calibration intuition (light) — reliability of predicted probabilities even within linear models; sets up Thursday without heavy math
  7. Transition to algorithmic methods + assign reproducibility mini-task (directory + render script)

Day 2 (Tuesday): Trees, rpart, and the tidymodels Pipeline

Learning goals

  • Compare mathematical and algorithmic modeling mindsets (two cultures)
  • Build intuition for decision trees (rpart) on gene simulations and Palmer Penguins
  • Master recipe → spec → workflow → tune → fit with decision_tree + rpart and accuracy on CV
  • Establish reproducible project layout (microbiome lab)

Sequence

  1. Breiman two cultures; gene rpart (classification + regression)
  2. Incremental pipeline walkthrough on penguins (day02-tidymodels-walkthrough.qmd — also Module 04)
  3. tune_grid for tree_depth and min_n; tuning plot, tree plot, light VIP
  4. Microbiome lab: same grammar on OTU data (grouped CV)

Day 4 (Thursday): Models, Preprocessing, Metrics, and Importance

Deck: day-04-thursday.html (includes: day04-*.qmd)

Learning goals

  • Swap parsnip specs (forest, boosting, MLP) on Tuesday’s rec_base
  • PCA, imputation, resampling as recipe steps; accuracy until imbalance
  • Metrics beyond accuracy, confusion matrices, ROC/PR on imbalanced data
  • VIP and non-causal interpretation (Module 06)

Sequence (matches slides)

  1. Part A: recap setup + stressed scenarios
  2. Part B: model catalog; bagging/boosting/MLP; multi-model fit_resamples (accuracy)
  3. Part C–D: step_pca(), step_impute_*() (accuracy)
  4. Part E: imbalance, metrics toolbox, confusion matrices, compare upsampling
  5. Part F: VIP on random forest; SHAP on penguin sex (no species) — game-theory intuition, beeswarm + waterfall, interpretation guardrails (Module 06)

Delivery Notes

  • Exercises are run separately from this hub but should reference Palmer Penguins (and optionally gene simulations) and reuse the data/model card templates.
  • Each day has one dedicated revealjs deck; optional live R snippets or board work replace separate app demos.
  • The same tidymodels pipeline grammar is reused from Day 2 (Tuesday) onward; Day 4 (Thursday) swaps engines and adds recipe steps (PCA, impute, upsample), not project structure.
  • caret removed from the hub; splits/scaling/metrics use tidymodels / yardstick in slides, modules, and notebooks.
  • Student .Rmd notebooks live under notebooks/: gene pair unchanged; penguin pair added; align exercises with the same penguin export where possible.

Optional extensions (only if time remains)

  • Deeper causal half-day (DAGs, backdoor adjustment) as a parallel track for quantitatively mature audiences.
  • Guest deployment expanded into a full lab with authentication and logging.
  • Versioned TA grading rubrics published to students day-one for every notebook hand-in.