Part IV
Language of Algorithms: Prediction, Segmentation, and Model Evaluation
From explaining the past to predicting the next move
This part turns the question forward — what will happen next, to whom, and what to do about it — and hands that work to algorithms operating at a scale no analyst could match by hand. The four chapters trace one arc: write the prediction problem down as a Task Contract and guard it against leakage (Chapter 9), build and grade supervised models on a threshold-profit curve rather than raw accuracy (Chapter 10), let the algorithm propose a lens when no target exists (Chapter 11), then push scores into a media budget through targeting, ranking, and the monitoring that keeps a live model honest (Chapter 12). The recurring discipline is that in an AutoML world the human leverage has migrated toward defining the task, crafting features, naming segments, and insisting on incrementality — because a high conversion rate proves selection, not causation.
4 chapters · 18 articles
What you’ll learn
- Write a leakage-proof Task Contract — target, features, unit, and label timing — and choose the right train/test split before any model is fit
- Grade a classifier on a manager's threshold-profit curve and read a model card instead of trusting raw accuracy
- Turn unlabeled behavior into named, action-ready segments using clustering and dimensionality reduction, with stability as the test
- Convert model scores into ranked, targetable audiences graded by precision@k and NDCG rather than a confusion matrix
- Keep a deployed model alive by separating data drift from concept drift and proving lift with an incrementality holdout
Chapters in this part
Get the task contract and the features right, and the algorithm almost picks itself; get them wrong, and no model can save you.
A model isn't evaluated until its scores meet the firm's cost matrix and ship with a card.
Unsupervised methods don't hand you answers — they hand you a lens, and a manager decides whether the structure is worth acting on.
Turning scores into audiences, rankings, and a monitored system that still works six months later.