Licence 3 MIASHS · semester 6 · R

Machine learning

Eight classification practicals on the Titanic data, then a group project: can you predict whether a fragrance is rated above the median from its accords and notes?

The project

Roughly 24 000 fragrances, a binary target, and five model families compared on ROC and confusion matrices, one of them rejected on stated grounds.

Test-set performance of the four retained models. Source: Projet/results.json, transcribed from the rendered report.

Read the ceiling, not just the ranking

The random forest wins, but every AUC sits between 0.65 and 0.69. The models separate the classes only modestly, which is the honest finding: accords and notes alone do not explain how people rate a perfume. Naive Bayes was dropped for a different reason, that its sensitivity collapses to 0.20 while its specificity climbs to 0.92, which is to say it wins by almost never predicting the positive class.

Read it The report 24 pages as PDF, or a navigable HTML book with the R folded into each chapter. Rendered Classification practicals Eight worked practicals on the Titanic data. Source The notebooks Seven Quarto chapters: exploration, feature engineering, one per model, comparison.

ROC curves are shown as figures inside the report rather than redrawn here: the curve points live in the model cache, which is rebuildable and not committed.