Scenarios 06–10
6 — Spam labels available
Situation: millions of emails already labeled spam/not spam; classify new messages.
Primary: learning-method selection with known labels.
Best: supervised learning; train/validate on representative data.
Weaker: unsupervised discovery ignores the known target.
Changed: labels disappear and goal is natural groups → unsupervised.
Source: pp. 480–481.
7 — Unknown customer segments
Situation: Marketing has no labels; wants natural groups from behavior.
Primary: hidden-pattern/group discovery.
Best: unsupervised clustering, then interpret/validate business meaning.
Weaker: supervised needs an authoritative target.
Changed: Marketing supplies authoritative segment labels and wants classification → supervised.
Source: pp. 480–482.
8 — Goal-feedback learner
Situation: game strategy improves actions from feedback about progress toward winning.
Primary: goal-directed feedback learning.
Best: reinforcement learning.
Weaker: supervised would require labeled example outcomes/moves.
Changed: complete labeled set of correct moves/outcomes is supplied → supervised may fit.
Source: pp. 480–481.
9 — Churn probability only
Situation: model produces next-month churn probability, no action recommendation.
Primary: predictive analytics.
Best: validate predictive performance and communicate uncertainty/use limits.
Weaker: prescriptive overstates the output.
Changed: system recommends the retention offer expected to reduce churn → prescriptive.
Source: pp. 474–475, 481–482.
10 — Live sensor shutdown
Situation: factory streams sensor data; model continuously scores failure and stops equipment above threshold.
Primary: operational analytics using a predictive model.
Supporting: Storage/Operations, Security, DQ.
Roles: Data Scientist, Platform Architect, operations owner.
Best: low-latency/streaming analytics plus monitoring and production controls.
Weaker: nightly batch cannot meet response time.
Changed: operations can act only next morning with no immediate harm → batch may be sufficient and cheaper.
Source: pp. 482–483, 490–491.