DataslopeSupervised vs Unsupervised Learning
The first great fork in machine learning. With labeled answers, you learn a mapping from inputs to outputs (supervised). Without them, you search the data for hidden structure (unsupervised). We make the distinction concrete with code.
On the last page you learned that the type of your target y tells you
what kind of problem you have. The very first fork in that decision is also
the biggest one in all of machine learning, and it hinges on a single
question: do you have answers to learn from, or not?
If your data comes with labels — a known answer attached to each example — you are in supervised learning. If it does not — you have inputs but no answers — you are in unsupervised learning. These two worlds use different algorithms, ask different questions, and are even evaluated in completely different ways. Knowing which one you are in is the first thing to settle on any project, so this page makes the line crisp.
The deciding question: is there a y?
Picture the difference with a teacher analogy, because the names come straight from it.
Supervised learning is learning with an answer key. A teacher (the
labels) tells the model the right answer for every training example. The
model's job is to learn the mapping well enough to answer new questions
where the key is hidden. You had a y, and you are predicting it.
Unsupervised learning is learning without an answer key. There is no
teacher and no y — only the raw examples X. With nothing to predict, the
model instead looks for structure: which examples are similar, which form
natural groups, how the data is organized. You are not predicting a known
answer; you are discovering what is there.
That tree is worth memorizing — it organizes most of classical machine
learning. The top split is supervised vs. unsupervised (do you have y?).
Each branch then splits again: supervised into regression vs. classification
by the type of y, and unsupervised into clustering vs. dimensionality
reduction by the kind of structure you seek. The next page zooms into the
three task types at the leaves; this page is about the trunk.
The label is the dividing line
Everything flows from one fact: the presence or absence of labels. A labeled dataset (features and answers) enables supervised learning. An unlabeled dataset (features only) calls for unsupervised learning. When you size up a new problem, your very first question should be: do I have labels, and are they the thing I actually want to predict?
Supervised learning: learning a mapping X to y
In supervised learning the model learns a function from inputs to a known
output: y = f(X). You show it examples with their answers, it learns the
relationship, and then it predicts answers for inputs it has never seen. This
is the setting of every example so far in the course — iris species, wine
classes, house prices — because supervised learning is the most common and
most directly useful kind.
Supervised learning splits in two by the type of y:
- Regression predicts a continuous number (price, temperature, demand).
- Classification predicts a category (spam or not, which species, which of five products).
Here is a compact supervised example. We train a LogisticRegression
classifier on the iris flowers — a supervised task, because every training
flower comes labeled with its true species — and then score it on held-out
flowers. The crucial detail: we pass both X and y to fit, because
supervised learning needs the answers to learn from.
Two signatures mark this as supervised. First, fit(X_train, y_train)
receives the answers. Second, we can compute an accuracy, because we have
the true y_test to compare predictions against. Both are only possible
because labels exist.
How to recognize supervised code at a glance
If you see fit(X, y) — two arguments, features and a target — it is
supervised. If you can later compute a metric like accuracy or error by
comparing predictions to known answers, it is supervised. The presence of y
on both ends (training and evaluation) is the tell.
Unsupervised learning: finding structure with no y
Now remove the answer key. In unsupervised learning you have only X, and
the model's job is to reveal structure that was already latent in the data.
The most common form is clustering: grouping examples so that members of
a group are similar to each other and different from other groups — without
anyone ever telling the model what the "right" groups are.
We will cluster synthetic data made of three natural blobs of points.
Critically, although make_blobs can hand back the true blob labels, we
throw them away and never give them to the model. KMeans sees only the
point coordinates X and must discover the groupings on its own.
Look at how different the call is. fit(X) takes one argument — there is
no y to pass. The labels that come out were created by the algorithm
("these points belong together"), not learned from any answer key. The model
found three groups because we asked for three and the data genuinely
clustered into three; it had no idea what those groups "mean."
Cluster labels are arbitrary names, not the truth
A clustering algorithm's group numbers (0, 1, 2) are just labels for groups it invented. They have no fixed meaning and might come out as 0, 1, 2 on one run and 2, 0, 1 on another — the grouping is the same, only the numbering differs. You cannot compute "accuracy" the way you do in supervised learning, because there is no ground-truth answer key to compare against. Evaluating clusters needs entirely different tools — covered in the evaluating clusters chapter.
Seeing them side by side
Putting the two calls next to each other makes the distinction unmistakable.
Same library, same fit method — but one gets answers and one does not.
The signature difference — fit(X, y) versus fit(X) — is the entire idea
in one line of code. If the algorithm needs answers, it is supervised. If it
works on features alone, it is unsupervised.
A brief word on dimensionality reduction
Clustering is the headline unsupervised task, but it is not the only one. Dimensionality reduction is another: it takes data with many features and compresses it into far fewer, while preserving as much of the important structure as possible. It is used to visualize high-dimensional data in 2-D and to denoise or simplify inputs before modeling. We mention it here only so the map is complete; it has its place in the broader unsupervised family alongside clustering.
The hidden reason this distinction matters: labels are expensive
There is a practical force lurking behind the supervised/unsupervised divide that the textbook framing often skips: labels cost money and time to obtain. Features are frequently cheap — every click, transaction, or sensor reading is logged automatically. But the answer attached to each example often has to be supplied by a human: a doctor must read each scan, a reviewer must mark each email as spam, an analyst must confirm each transaction was fraud. That labeling effort can dwarf every other cost in a project.
This is why the choice is not purely "which question am I asking?" but also "what data can I realistically get?" You might want a supervised model, yet have only a mountain of unlabeled data and no budget to label it. In that situation, unsupervised methods let you extract value now — finding structure, surfacing anomalies, organizing the data — while you decide whether labeling is worth it.
A useful middle ground: semi-supervised learning
Reality is not always all-or-nothing. Sometimes you can afford to label a small fraction of your data but not all of it. Semi-supervised learning combines a little labeled data with a lot of unlabeled data, using the structure in the unlabeled portion to make the few labels go further. You do not need the details now — just know that the labeled/unlabeled split is a spectrum, not a hard wall, and that labeling effort is often the real constraint that decides which approach is feasible.
Two worlds, two ways to evaluate
To cement how differently these two settings behave, look at evaluation side
by side in code. The supervised model can be scored against known answers; the
clustering result cannot — there is simply no y_true to compare to. The
contrast in what is even possible is the clearest way to feel the divide.
The supervised side reports accuracy because there are true labels to check against. The unsupervised side has no such luxury, so it falls back on an internal measure — here the silhouette score, which rewards tight, well-separated groups without needing any ground truth. You do not need to understand the silhouette score yet (the evaluating clusters chapter handles it); the point is simply that the two settings demand entirely different evaluation toolkits, a direct consequence of one having answers and the other not.
Do not import supervised habits wholesale
Coming from supervised learning, it is tempting to expect a single "score" for everything. Unsupervised learning breaks that expectation: there is no one number that says a clustering is "right," because rightness is not even defined without labels. Judging unsupervised results leans on internal measures and human judgment about whether the structure is useful — a very different mindset from chasing a test-set accuracy.
A quick check
What single fact most fundamentally separates supervised from unsupervised learning?
Supervised learning uses scikit-learn; unsupervised learning does not
Supervised learning is for numbers; unsupervised learning is for text
Supervised learning has labeled answers (a target y) to learn from; unsupervised learning has only features X and must find structure without answers
Supervised learning is always more accurate
When each is appropriate
Reach for supervised learning when you have a clear, specific answer you want to predict, and you have historical examples where that answer is known. Predicting whether a customer will churn? If you have past customers labeled "churned" or "stayed," that is supervised. Forecasting next quarter's revenue from past quarters? Supervised. The hallmark is a well-defined target plus labeled history to learn from.
Reach for unsupervised learning when you have no labels — or when your question is exploratory rather than predictive. You are not asking "what is the answer for this example?" but "what natural structure exists in my data?" Segmenting customers into groups when you have no pre-defined segments? Unsupervised. Exploring a fresh dataset to see whether it falls into natural clusters before you even know what to predict? Unsupervised.
A practical sequence
The two are often used together, in order. Faced with a brand-new dataset and no labels, you might first run unsupervised clustering to understand its structure and generate hypotheses. Then, once you have labels (perhaps by having experts label a sample), you switch to supervised learning to build a predictive model. Exploration first, prediction second.
Why you cannot "just score" unsupervised learning
This trips up almost everyone coming from supervised learning, so it is worth stating plainly. In supervised learning, evaluation is conceptually simple: the model predicts, you compare to the known answers, you compute accuracy or error. The answer key makes scoring straightforward.
In unsupervised learning there is no answer key. If a clustering algorithm splits your customers into four groups, there is no "true" grouping written down anywhere to check it against — the whole reason you clustered is that you did not know the groups. So "accuracy" is undefined. Instead, unsupervised methods are judged by internal measures — how tight and well-separated the discovered groups are — and, ultimately, by whether the structure is useful to a human. The evaluating clusters chapter develops this carefully; for now, just absorb that supervised and unsupervised learning are evaluated in fundamentally different ways, and you cannot port accuracy across the line.
The most common cross-over mistake
Do not try to compute supervised metrics (accuracy, precision, error) on a pure clustering result as if the invented cluster numbers were predictions of some true label. Without ground truth, those numbers are meaningless. If you do happen to have true labels lying around, you are no longer doing pure unsupervised learning — and you might reconsider whether a supervised approach fits your goal better.
Common misconceptions
- "Unsupervised learning is just supervised learning without the labels filled in." No — the goal is different. Supervised learning predicts a known answer; unsupervised learning discovers unknown structure. They are not two settings of one task; they answer different questions.
- "Unsupervised learning needs no human judgment." The opposite is often true. With no answer key, deciding whether the discovered structure is meaningful — and how many clusters to even ask for — leans heavily on domain knowledge and human interpretation.
- "More clusters always means a better clustering." Push the count high enough and every point becomes its own group — technically "tight," utterly useless. Choosing a sensible number of clusters is a real and subtle decision, not a maximization.
- "Supervised is better than unsupervised." Neither is better; they solve different problems. The right choice is dictated by whether you have labels and what question you are asking, not by a ranking.
Real-world applications
Supervised powers most of the predictive systems you interact with daily: spam filters (labeled spam / not-spam), medical diagnosis from labeled scans, credit scoring from labeled defaults, demand forecasting from labeled historical sales, image recognition from labeled photos. Wherever there is a specific answer to predict and labeled history to learn from, supervised learning is the workhorse.
Unsupervised shines in discovery and exploration: customer segmentation (grouping shoppers with no predefined segments), anomaly detection (flagging points that fit no normal group, useful in fraud and equipment monitoring), organizing large document or image collections by similarity, and compressing high-dimensional data for visualization. Wherever you want to understand the shape of data you do not yet have labels for, unsupervised learning leads.
Your turn
The challenge below tests the judgment this whole page builds: given a described situation, decide whether it calls for supervised or unsupervised learning by asking the one deciding question — is there a labeled target to predict?
For each scenario, decide whether it is a supervised or unsupervised learning problem, using the deciding question: does each example come with a known label/answer you want to predict?
Fill in the dictionary answers so each scenario key maps to either the
string "supervised" or the string "unsupervised":
"predict_house_price"— predict a house's sale price using a dataset of past houses where the actual sale price is known for each."segment_customers"— group shoppers into segments when you have NO predefined segments, only their purchase behavior."classify_email_spam"— label new emails as spam or not, trained on emails already labeled spam / not-spam."group_news_articles"— organize a pile of unlabeled news articles into natural topic groups discovered from the text."diagnose_from_labeled_scans"— predict a diagnosis from medical images that experts have already labeled with the correct diagnosis.
The hidden tests check each individual answer.
Check your understanding
In scikit-learn, which call signature signals a supervised algorithm?
fit(X) — features only
fit() — no arguments at all
fit(X, y) — features and a target, because supervised learning needs the answers to learn the mapping
fit(y) — the target only
You have a dataset of shoppers' purchase histories but no predefined customer segments, and you want to discover natural groups. What kind of learning is this?
Supervised regression, because purchases are numbers
Supervised classification, because shoppers belong to types
Unsupervised learning (clustering), because there are no labels — you are discovering structure, not predicting a known answer
It is not a machine learning problem at all
Why can't you compute ordinary "accuracy" for a pure clustering result?
Because clustering algorithms are too slow to evaluate
Because accuracy only works for regression, not for groups
Because there is no ground-truth answer key — the cluster numbers were invented by the algorithm, so there is nothing "correct" to compare them against
Because clusters always achieve 100% accuracy by definition
A team has a large set of customer records and wants to (a) first understand whether the data falls into natural groups, then (b) later predict churn once they have labels. Which sequence of approaches fits?
Supervised first to explore, then unsupervised to predict
Unsupervised for both steps, since labels are never needed
Unsupervised first to discover structure with no labels, then supervised to predict churn once labels are available
Neither approach applies to customer data
Which statement is true about the relationship between supervised and unsupervised learning?
Unsupervised learning is just supervised learning with the labels hidden, solving the same task
Supervised learning is always the better choice when both are possible
They solve fundamentally different problems — predicting a known answer versus discovering unknown structure — and are evaluated in completely different ways
Both always require a target column y to function
Features and Targets
Every supervised machine learning problem reduces to a feature matrix X and a target y. We define both precisely, walk through the types of features and targets, and practice splitting a real table into the two pieces every model expects.
Regression, Classification, and Clustering
The three core task types of classical machine learning, told apart by one thing — what comes out. A number means regression, a category means classification, and groups with no labels mean clustering. We see all three in code and pictures.