ml

claude/business_model.md

@@ -0,0 +1,43 @@
+Yes. Most business questions map to a model family once you work out the unit of analysis (who or what is being counted) and the type of outcome. Here's the mapping using your Trakt and Spotify examples.
+
+**Who**
+- *Questions:* Who are my heavy skippers? Who is most like whom?
+- *Frameworks:* clustering (K-means) for segments, logistic models for propensity (e.g. who completes shows), similarity scores (Jaccard, cosine) for how alike two users are.
+- *Example:* Spotify user A vs. B overlap, or the viewing-day personas from K-means.
+
+**What**
+- *Questions:* What should we recommend? What drives completion?
+- *Frameworks:* recommenders (item-item, Word2Vec), association rules (FP-growth), and GLM coefficients that rank drivers.
+- *Example:* tracks B plays that A hasn't heard, or runtime's effect on finishing a show.
+
+**When**
+- *Questions:* When will demand peak? How long until someone abandons a show? When did habits change?
+- *Frameworks:* time series (ETS, Prophet) for forecasts, survival analysis for time-to-event, Poisson with seasonality terms, change-point detection.
+- *Example:* predicted days until a 50% chance of abandoning a series.
+
+**Where**
+- *Questions:* Which region, device or channel?
+- *Frameworks:* group-level GLMs with country or platform terms, hierarchical models to shrink small-sample regions toward the average.
+- *Example:* Spotify's country and platform fields. Trakt has almost no "where", so there it's mostly a time-zone question.
+
+**How many / how much**
+- *Frameworks:* Poisson or negative binomial for counts, gamma for skewed amounts, with an offset for exposure.
+- *Example:* plays per day on weekends vs. weekdays.
+
+**Why**
+- *Frameworks:* GLM coefficients show association. Causal designs (A/B tests, difference-in-differences) show effect.
+- *Example:* "Did B hearing it first cause A's plays?" needs more than a coefficient.
+
+**The shortcut:** ask what the outcome looks like.
+- A count → Poisson
+- Yes/no → logistic
+- Time until something → survival
+- A skewed amount → gamma
+- Groups of similar things → clustering
+- Next best item → recommender
+
+**Two cautions**
+- Plain SQL answers many who/what/when/where questions descriptively (top artists, busiest days). Reach for a model when you need prediction, comparison with uncertainty, or an effect size.
+- Models give probabilities and rates, not certainties, and sparse data (one or two users) limits which ones are valid.
+
+Want to take one real business question from your data and walk it through question → outcome type → model → features?