Classification

Classification is the machine-learning task of putting things into named boxes that you defined ahead of time. You feed an algorithm a pile of past examples that already carry their correct labels — emails marked spam or safe, customers marked churned or retained, transactions marked fraud or legitimate — and it learns the patterns that separate one box from another. Then, faced with a new record it has never seen, it predicts which box that record belongs in. Because the algorithm learns from answers you supplied, classification is a form of supervised learning. The boxes can be two (binary classification, like buy or no-buy) or many (multiclass, like sorting support tickets into billing, technical, or account). What unites them all is that the set of categories is fixed and known before training begins.

For marketers, classification is the quiet engine behind a great deal of automation. Lead scoring is classification: the model studies which past leads converted and which did not, then labels each new lead as likely or unlikely to buy, so sales can prioritize. Churn prediction is classification: it learns the signatures of customers who left and flags the ones at risk while there is still time to intervene. Spam filtering, sentiment tagging of reviews, image moderation, and fraud screening are all the same task in different clothes. The value is that it turns a vague human judgment — "this feels like a hot lead" — into a repeatable, scalable decision backed by evidence from thousands of prior cases, and it produces not just a label but often a probability you can rank and threshold.

Classification is most often confused with clustering, and the difference is fundamental. Classification is supervised — it needs labeled examples and predicts a category you already defined, like "high-value" or "low-value." Clustering is unsupervised — it gets no labels and instead discovers natural groupings in the data on its own, then you decide what those groups mean afterward. Put plainly, classification sorts records into buckets you named in advance, while clustering invents the buckets. If you already know the customer tiers you care about and have past examples of each, you classify. If you suspect there are meaningful segments but cannot name them yet, you cluster, often using the result to create labels you then classify against.

Classification also differs from regression, its sibling in supervised learning. Regression predicts a number on a continuous scale — next month's revenue, a customer's expected lifetime value, a likely order size. Classification predicts a category from a discrete set — will this customer churn, yes or no. The line can blur, because many classifiers output a probability between zero and one, but the decision they support is categorical: cross a threshold and the lead is "hot." Choosing between them comes down to the question. "How much will they spend" is regression; "will they buy at all" is classification. Common classification algorithms include logistic regression, decision trees, random forests, gradient boosting, and neural networks, each learning the boundary between classes in its own way.

Use classification well by judging it on the right metrics, not just raw accuracy. Accuracy alone is treacherous when classes are imbalanced — if only two percent of leads convert, a model that predicts "no" for everyone is ninety-eight percent accurate and completely useless. Lean instead on precision (of the leads you flagged hot, how many really were), recall (of the truly hot leads, how many you caught), and the confusion matrix that lays both bare. Decide which error costs more: missing a churning customer or wasting a save offer on a loyal one, then tune the decision threshold to match that cost. Watch for leakage, where a feature secretly encodes the answer, which makes a model look brilliant in testing and fail in production.

Treat the labels themselves as the foundation, because a classifier is only as honest as the examples it learned from. If your historical "converted" labels are sloppy or biased, the model inherits that. Retrain as behavior drifts, since last year's signature of a hot lead may not hold this year. Pair classification with segmentation and incrementality thinking so you act on the predictions rather than just admire them. The failure mode is shipping a high-accuracy number that hides a model blind to the rare class you actually care about — the fraud, the churn, the conversion — and acting on its labels without ever checking the costs of being wrong.

Classification, from the Latin for "arranging into classes," is the supervised task of assigning records to predefined categories learned from labeled examples.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where classification is a core concern: