🤖 AI / ML

Supervised Learning

supervised learning · supervised machine learning · labelled learning · labeled training data · classification and regression

Definition

Supervised learning is a machine learning paradigm in which a model is optimised on labelled data, where every input example carries a correct target answer — a class or a number.

How it works

In supervised learning the model receives a dataset of pairs: an input feature vector X and a
target value y. The task is to learn a function f(X) → y that minimises prediction error on new
data.

In e-commerce that looks like this:
– X — purchase history, views, demographics, time since the last purchase
– y — the purchase itself (1/0), churn probability (0–1), the expected order value

Input (features): [7 purchases in 90 days, last one 14 days ago, 3 categories, AOV $25]
Target value:     churn = 0  (did not leave within the next 30 days)

The two main task types

Classification — predicting a categorical answer (yes/no, class A/B/C). Examples: will buy or
will not buy, will churn or will stay, search intent transactional or informational.

Regression — predicting a numeric value. Examples: expected LTV, the forecast order value of the
next purchase, the probability of a return.

Use in personalization

Recommendation algorithms built on supervised learning are trained to predict the probability of an
interaction — a click, a purchase — for a user-item pair:

Task	Features (X)	Label (y)
Recommendations	User profile plus product attributes	Click or purchase
Churn prediction	RFM features plus behaviour	Churn within the next 30 days
PLP ranking	User plus position plus product	CTR or CR

A critical dependency on data

Label quality determines model quality. The typical problems in e-commerce:

Sample bias: the model is trained only on purchased products and never sees the items a
shopper viewed and abandoned because the page was poor
Data leakage: features accidentally contain data from after the target event
Class imbalance: a purchase happens in 2% to 3% of cases, so the model takes the lazy route
and stops predicting the rare class

Frequently asked questions

Which e-commerce tasks are solved with supervised learning?

Purchase probability prediction, search intent classification (informational versus transactional query), churn prediction, return probability estimation, and listing ranking by the likelihood of a click or a purchase.

How is it different from unsupervised learning?

In unsupervised learning the data carries no labels and the model looks for structure on its own — clusters, patterns. An example is segmenting shoppers by behaviour with no categories defined in advance. Supervised learning always needs a labelled dataset with correct answers.

What is a target label and where does it come from in e-commerce?

The label is the fact the model has to predict. For churn prediction it is 1 (churned) or 0 (stayed). For recommendations it is a click, an add-to-basket or a purchase. Historical transaction data is the main source of labels for most e-commerce tasks.