Targeting, Ranking, and Operating ML

This chapter focuses on the point where algorithmic ideas stop living in a notebook and start spending a media budget. It follows the bridge from a clustered segmentation to a targetable audience on an ad platform, reframes a lookalike audience as nearest-neighbour scoring run at platform scale, then turns scores into ranked lists through collaborative, content-based, and learned recommenders graded with precision@k and NDCG. It closes the lifecycle with the discipline that keeps a model alive — separating data drift from concept drift, designing a four-KPI monitoring dashboard, setting retraining cadences, and weaving everything into a single Bean & Basket Customer Intelligence loop. The recurring lesson: high conversion proves selection, not causation, and only an incrementality holdout separates real lift from customers who would have converted anyway.