Framing: classification, regression, clustering, forecasting
Lesson 1 of 5 in Choosing a Modeling Approach.
Nearly every Domain 2 scenario on MLA-C02 starts the same way: "A company wants to…" Before any service name matters, the blueprint wants one decision made correctly — what kind of problem is this? Translate the business verb into an ML problem family, and half the answer options eliminate themselves.
The families you must recognize on sight: classification (which bucket?), regression (what number?), clustering (what groups exist, with no labels?), forecasting (what happens next in a time series?), and recommendation (what will this user want?). Anomaly detection rides along as a sixth — "find the weird ones."
| Family | The question | Output | Representative AWS route |
|---|---|---|---|
Classification | Which category? | A label (+ probability) | XGBoost / Linear Learner built-ins; Comprehend for text |
Regression | How much? What value? | A continuous number | XGBoost / Linear Learner in regression mode |
Clustering | What natural groups exist? | Group assignments (no labels needed) | K-Means built-in |
Forecasting | What comes next over time? | Future values with uncertainty | DeepAR built-in (many related series) |
Recommendation | What will this user engage with? | Ranked items per user | Amazon Personalize (pre-trained service) |
Anomaly detection | Which points are abnormal? | Anomaly scores | Random Cut Forest built-in |
Interactive sorting exercise: Frame each business request as the right ML problem family.
Two quick framing reflexes finish the job. Labels present? If historical outcomes exist (churned / didn't, sale price), it's supervised — classification or regression by output type. Words like segment, group, discover with no labels mean unsupervised. Time in the sentence? "Next month," "per day," "over the holiday season" over an ordered history means forecasting, not plain regression. And recommendation is usually not a build decision at all — Amazon Personalize ships it as a service.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.