| name | framing-ml-problems |
| description | Frames a data/ML problem before any modeling — the decision, the action, the cost of a wrong call, task type, target, and success metric. Use when picking a project direction, writing a research question, or mapping a problem onto an ML task type. |
Framing ML problems
A model is never the goal. A better DECISION is the goal. Frame first, model later.
The four questions (answer all four, in writing)
- What decision does this improve? Not "predict X" — decisions. "Which page should an
editor fix first?" is a decision. "Predict decline" is not, yet.
- Who acts on the output, and what do they do? Name the person and the action. If nobody
would act differently, the work has no customer.
- What does a wrong answer cost? Wasted editor hours? A missed decline? The cost of errors
decides how careful the method must be, and which errors matter more.
- Why does data or ML help at all? Sometimes a plain rule (an if-statement) is the right
answer. Sometimes a dashboard is. ML earns its place only when the pattern is real but too
messy to write by hand — many signals, tangled, shifting over time.
Map it to a task type
| If your question sounds like… | Task type | Target | Typical metric |
|---|
| "Which ones first?" | Ranking / scoring | a priority score | precision@K |
| "Will this one decline / recover?" | Classification | a yes/no label from an OBSERVED outcome | ROC-AUC, precision/recall vs base rate |
| "What kinds of items exist?" | Clustering | none (unsupervised) | silhouette + human sense-check |
| "Which signals travel together?" | Signal analysis | none | effect sizes, grouped comparisons |
Two rules that save whole projects:
- The target must be observed, not defined. A label that comes from someone's rule means your
model learns the rule, not the world. Prefer outcomes measured in a later time window.
- Name the metric before training. "Good" defined after the fact always looks good.
Write the one-paragraph frame
For [who], deciding [what], we will build [output type] from [data], predicting/scoring
[target] measured by [metric]. A wrong call costs [cost]. A plain rule isn't enough because
[why]. We will claim only [observed / directional / decision-support] results.
If you can fill that paragraph honestly, you are framed. If a blank stays blank, that blank IS
your next task.
How to verify
- Show your frame to someone (or your assistant) and ask: "what decision does this improve?"
If their answer differs from yours, the frame isn't done.
- Check the target: is it observed in the data, or defined by a rule? If defined — reframe.
- Check the metric exists in your data. Can you compute it today, on a baseline? Do it.
