Exam-style
A marketing team has three years of subscriber behaviour: box sizes, delivery frequency, substitutions accepted, pauses taken, spend per month. No column records which group a subscriber belongs to, and nobody has ever written such a group down. They want to know whether there are natural groups of subscribers worth treating differently. They have no idea in advance how many groups there might be, or what the groups would be called. Which technique fits?
Reveal the answer
C. Clustering, an unsupervised method that groups subscribers by similarity with no target column to predict
Two facts about the data settle this before any algorithm is named. There is no target column, so nothing exists for a model to be trained to predict. There is no fixed set of categories either, so nothing exists for a model to predict into. Both halves of supervised learning are missing, which leaves unsupervised learning, and the unsupervised method that proposes groups is clustering. Classification, in either its multi-class or its binary form, needs the categories defined in advance and needs labelled examples of each one; here the categories are the thing being looked for, so a classifier has nothing to learn from. Regression fails for the same reason. It predicts a number, which means a column of past numbers to learn against, and banding a predicted spend draws the groups along one axis somebody chose rather than along the shape the behaviour actually has. Semi-supervised learning is the tempting one, because it is the option that admits the labels are missing. It fits when the categories are already known and the labels are scarce. A few thousand hand-labelled records then teach the model enough to guess at the rest. Reach for it here and the labelling step decides the answer. Whoever labels the sample has to invent the segments to label with, which imposes the categories instead of discovering them, and every group the model finds afterwards is one of the groups that person made up. Clustering keeps the discovery in the data. Two caveats come with it. How many groups to look for is usually a setting a person picks and tries several values of, not an answer the data hands back. And the output is unnamed: reading the cluster centres and deciding that group three is the fortnightly bulk buyers is human work, done after the run.
Q. Three years of subscriber behaviour, no labels, no target column, and nobody knows how many groups there are or what to call them. Which technique?
A. Clustering, which is unsupervised learning: it groups records by similarity with nothing to predict.
Why? Supervised learning needs two things this data lacks. Classification needs a fixed set of labels agreed in advance and labelled examples of each; regression needs a column of past numbers to learn against. Data arrives labeled and unlabeled, and unlabelled data rules out both. Semi-supervised learning is the near miss. It fits when the categories are known and the labels are thin on the ground. Labelling a sample here means inventing the segments before you look for them, which answers a different question entirely.