Supervised vs Unsupervised Learning Explained (with Examples)

Two ways to learn a subject. One: a teacher gives you worked examples with the answers, you study the pattern, and then you attempt fresh questions. Two: nobody gives you answers at all — you are handed a pile of material and asked what it groups into.

Those are supervised and unsupervised learning, and the choice between them isn't yours. The data you have decides.

The question that decides

Do I have the right answer for each example? That single question routes almost every problem you will meet.

And one clarification that trips people up in interviews: "supervised" refers to the labels, not to a person watching. Nobody is supervising the training. The data carries its own supervision.

Supervised learning

Every training row comes with the correct answer, and the algorithm learns the mapping from inputs to that answer. It splits into two tasks by the type of the target:

# Classification — the target is a category
X = df[["amount", "hour", "distance_from_home", "prior_declines"]]
y = df["is_fraud"]              # 0 or 1, labelled from history

model = RandomForestClassifier().fit(X_train, y_train)
model.predict_proba(new_txn)    # -> probability of fraud

# Regression — the target is a number
X = df[["area_sqft", "bedrooms", "distance_metro_km"]]
y = df["rent"]                  # ₹, a continuous value

model = LinearRegression().fit(X_train, y_train)
model.predict(new_flat)         # -> a rupee amount

💡 Notice predict_proba in the classification example. Most classifiers can give you a probability, not just a class — and in production that is usually more useful, because it lets you rank cases and set a threshold from business economics (lesson 6).

Unsupervised learning

No labels. The algorithm finds structure in the data itself. Three common tasks:

• Clustering — grouping similar items. Customer segmentation, grouping news articles by topic when nobody has defined the topics.
• Dimensionality reduction — compressing many correlated features into a few, mostly for visualisation, noise reduction or speed (PCA, lesson 8).
• Anomaly detection — learning what normal looks like and flagging deviations, when you have no examples of "abnormal".

from sklearn.cluster import KMeans

X = df[["recency_days", "order_count", "total_spend"]]   # no y at all

model = KMeans(n_clusters=4, n_init=10).fit(X)
df["segment"] = model.labels_

print(df.groupby("segment")[["recency_days", "order_count", "total_spend"]].mean())

The algorithm found four groups. It did not tell you what they mean — that reading is yours: segment 0 looks like loyal regulars, segment 1 like lapsed customers, segment 2 like infrequent big spenders. Unsupervised learning produces groups, not meanings.

Reinforcement learning

A third mode, and one you should be able to describe even though you're unlikely to build one at placement level. An agent takes actions in an environment and receives rewards or penalties. Over many attempts it learns which sequences of actions lead to the best outcome.

No labelled "correct move" exists — only the consequence. Game playing, robotics, and sequencing decisions are the classic applications. It also appears in LLM training as RLHF, which is why the term shows up in AI interviews (fine-tuning lesson 2).

Why it is rare in business settings: it needs a safe place to experiment and a clear numeric reward. Neither is easy when the experiments involve real customers and real money.

Wait — how do you grade something with no answers?

This is the honest difficulty of unsupervised learning, and a good interview question.

With labels, evaluation is straightforward: compare predictions to the truth. Without them, there is no ground truth to compare against. You can measure internal properties — the silhouette score tells you whether clusters are tight and well-separated — but that measures geometry, not usefulness.

So the real evaluation is a judgement call: are these groups meaningful and actionable to the people who will use them? Six mathematically tidy customer segments that marketing cannot run different campaigns for are worse than four they can.

🎯 Selection-round radar: "Difference between supervised and unsupervised learning" is one of the most-asked ML questions anywhere. Define both, give one example each, and then add the two lines that separate you: classification versus regression splits supervised by target type, and unsupervised results have no ground truth, so a human has to decide whether the output is useful.

The realistic middle ground

Real projects rarely have either perfect labels or none. Two terms worth knowing:

Semi-supervised learning — a small labelled set and a large unlabelled one. Common when labelling is expensive: 200 X-rays read by a radiologist and 10,000 unread.

Transfer learning — start from a model already trained on a large general dataset and adapt it to your small one. For the X-ray case this is usually the strongest practical answer, and it is the same idea as fine-tuning an LLM (fine-tuning lesson 1).

There is also self-supervised learning, where labels are generated from the data itself — predicting the next word in a sentence needs no human labeller, which is exactly how large language models are pretrained.

Common mistakes

• Thinking "supervised" means a human watches training.
• Calling a numeric prediction "classification" or vice versa.
• Expecting a clustering algorithm to name or explain its groups.
• Choosing a learning type by preference rather than by the data available.
• Discarding unlabelled data instead of using transfer or semi-supervised learning.
• Reporting a silhouette score as if it proved the clusters were useful.

Quick recap

Practice Zone — PYQs from real selection rounds

Six MCQs, then two tasks: route six problems to the right learning type, and explain all three to a non-technical manager.

Q1Types of learningeasy

What defines supervised learning?

Asked in

aA human watches the training processbThe training data includes the correct answer (label) for each example, so the model learns to map inputs to known outputscThe model supervises itselfdIt only works on images

Q2Types of learningeasy

Which is an unsupervised learning task?

Asked in

aPredicting house prices from past salesbGrouping customers into segments when nobody has defined the segments in advancecClassifying email as spam using labelled examplesdPredicting whether a loan will default

Q3Types of learningmedium

What is the difference between classification and regression?

Asked in

aClassification is fasterbClassification predicts a category from a fixed set; regression predicts a continuous numbercRegression only works on imagesdClassification needs no training data

Q4Types of learningmedium

What is reinforcement learning?

Asked in

aTraining a model twice for reinforcementbAn agent learns by acting in an environment and receiving rewards or penalties, discovering which sequences of actions lead to the best outcomecAdding more data to a supervised modeldA type of unsupervised clustering

Q5Types of learninghard

You have 10,000 medical images but only 200 are labelled by a doctor. What approach fits best?

Asked in

aDiscard the 9,800 unlabelled imagesbSemi-supervised learning — use the labelled set to train and the unlabelled set to learn general structure, or use transfer learning from a pretrained model and fine-tune on the 200cUse unsupervised learning only and hope the clusters match the diagnosesdLabel the rest randomly

Q6Types of learningmedium

Why is unsupervised learning harder to evaluate than supervised learning?

Asked in

aIt runs more slowlybThere is no ground truth to compare against — you can measure internal properties like cluster tightness, but whether the result is useful is a judgement callcIt uses more memorydIt cannot be evaluated at all

Hands-on tasks:

FAQ

Next lesson: the first and most important supervised algorithm — Lesson 3: Regression Explained →