DataslopeRegression, 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.
By now you can ask the two questions that frame any machine learning problem: do I have labels? (supervised vs. unsupervised) and what type is my target? (number vs. category). Combine those two questions and you land on one of three task types that cover the overwhelming majority of classical machine learning:
- Regression — predict a number.
- Classification — predict a category.
- Clustering — discover groups, with no labels at all.
The fastest way to tell them apart is also the most reliable: look at what the model outputs. Out comes a number on a continuous scale? Regression. Out comes a label from a fixed set of classes? Classification. Out come group assignments invented from unlabeled data? Clustering. This page makes each one concrete — in code, and in a picture — so the three become instantly recognizable.
Tell them apart by the output, every time
When you are unsure which task you face, ignore the algorithm name and ask one question: what does a single prediction look like? A dollar amount or a temperature is regression. A label like "spam" or "setosa" is classification. A group id assigned to unlabeled data is clustering. The shape of the output is the surest classifier of the task.
Regression: predicting a number
Regression predicts a continuous quantity — a value on a scale where in-between answers make sense. House prices, temperatures, a patient's blood pressure, tomorrow's demand: all numbers that could land anywhere in a range, not just on a few fixed options.
The mental image is a line (or curve) through a cloud of points. The
model learns the trend relating the input to the output, and to predict, it
reads the height of that trend at a new input. Let us train the simplest
regressor, LinearRegression, on synthetic data with one feature, and draw
the line it learns.
The output is a number, and it is continuous: nudge the input and the prediction slides smoothly. There is no notion of "categories" here — the model is estimating a quantity. That is the unmistakable signature of regression. We will measure how good such predictions are (with metrics like mean absolute error and R²) in the regression metrics chapter; here, focus on the shape of the task: number in, number out.
Is it really continuous?
A quick test for regression: could the true answer reasonably be a value between two of your observed answers? Between a price of 300,000 and 310,000 sits 305,000, a perfectly sensible price — so price is continuous, and predicting it is regression. If "in-between" answers are nonsensical (half-way between "cat" and "dog"?), you are not looking at regression.
Classification: predicting a category
Classification predicts a category — one label from a fixed, finite set. Spam or not spam (two classes); the species of an iris (three classes); which of five products a customer will buy (five classes). The output is not a quantity but a choice among options, and "in-between" answers do not exist: an email is spam or it is not.
The mental image is a boundary that carves up the feature space into
regions, one region per class. To classify a new point, the model sees
which region it falls in. We will train LogisticRegression on two iris
features (so we can plot it in 2-D) and color each flower by its species.
The output is a label, drawn from a fixed set of three species. There is no "2.5th species." That discreteness — a choice among named classes — is what makes this classification rather than regression, even though the inputs (petal measurements) are themselves continuous numbers. Remember: the task type is defined by the output, not the inputs.
Numbers used as class labels are still categories
A frequent confusion: iris classes are stored as 0, 1, 2 — numbers! — so people assume the task is regression. It is not. Those integers are just names for categories; class 2 is not "twice" class 1, and there is no class 1.5. Whenever the target is a fixed set of labels — whatever symbols encode them — the task is classification. The math underneath treats them as distinct categories, not as quantities on a scale.
A quick check
A model outputs a person's predicted annual income in dollars — any value on a continuous scale. What task type is this?
Classification, because income comes in brackets
Clustering, because incomes form groups
Regression, because the output is a continuous number that could land anywhere on a scale
It cannot be determined from the output alone
Clustering: discovering groups with no labels
Clustering is the odd one out: it is unsupervised, so there are no labels at all. Instead of predicting a known answer, it discovers groups — sets of examples that are similar to each other and different from the rest. Nobody tells the algorithm what the groups are or even what they mean; it finds them from the structure of the features alone.
The mental image is points falling into natural clumps, with the
algorithm drawing a circle around each clump. We will run KMeans on blobs,
asking it to find three groups, and color the points by the group it
assigns. Notice we never pass any labels.
The colors here did not come from any answer key — they are invented by the algorithm. Re-run with a different seed and the same three clumps might get numbered differently; the grouping is what matters, not the labels 0/1/2. Because there is no ground truth, you cannot compute "accuracy"; clustering is judged by how tight and separated the groups are, and by whether they are useful — the subject of the evaluating clusters chapter.
Clustering does not know what the groups mean
A clustering algorithm finds that points group together, never why or what the groups represent. If you cluster customers and get three groups, the algorithm cannot tell you "these are bargain hunters, these are loyalists, these are one-time buyers" — interpreting the clusters is a human job requiring domain knowledge. The algorithm supplies structure; you supply meaning.
The three side by side
Here is the whole page in one table-like view. Same library, three different shapes of problem and output.
And the code signatures, distilled:
| Task | Has labels? | A single prediction is | scikit-learn call |
|---|---|---|---|
| Regression | Yes (a number) | a continuous value | reg.fit(X, y); reg.predict(X_new) |
| Classification | Yes (a class) | a category label | clf.fit(X, y); clf.predict(X_new) |
| Clustering | No | a discovered group id | km.fit(X); km.labels_ |
The reliable test
To name any task: (1) Are there labels? No labels and you are clustering. (2) If labeled, is the target a number or a category? Number is regression, category is classification. Two questions, three tasks, every time.
One dataset, three framings
The claim that "the task is set by the question, not the data" sounds abstract, so let us prove it on a single dataset. We will take the diabetes dataset — whose target is a continuous measure of disease progression — and frame the same data as all three task types, just by changing the question we ask. This is the most direct way to see that the task type lives in your intent, not in the rows.
Nothing about the underlying measurements changed between the three blocks. What changed was the question: "what is the exact score?" (regression), "is the score high or low?" (classification, after bucketing the target), and "which patients resemble each other?" (clustering, ignoring the target entirely). Same rows, three tasks — because the task type is a property of what you are trying to learn, not of the data sitting on disk.
Bucketing trades precision for simplicity
Turning a continuous target into categories (regression → classification by "bucketing") is a real and common choice, not a trick. You might do it because a coarse "high / medium / low risk" decision is all the business needs, or because classes are easier to act on. The cost is thrown-away precision: you no longer distinguish a score of 81 from 250 if both land in "high." Whether that trade is worth it depends entirely on how the prediction will be used — a recurring theme in framing problems well.
Common misconceptions
- "The inputs decide the task." They do not — the output does. A model with continuous numeric inputs can be a classifier (iris petals → species); a model with categorical inputs can be a regressor (neighborhood → price). Look at what comes out.
- "Integer labels mean regression." No. Class labels are often stored as integers (0, 1, 2) for convenience, but they are categories, not quantities. If there is no meaningful "in-between," it is classification.
- "Clustering is just classification without training the labels." No — there are no labels at all, and there is no ground truth to be right or wrong against. The goal is discovery, not prediction of a known answer.
- "You must pick exactly one task per dataset." A single dataset can support several. Customer data could feed a regression (predict spend), a classification (predict churn), and a clustering (find segments). The task is set by the question you ask, not by the data alone.
- "More clusters is always better." Crank
n_clustersto the number of points and every point is its own perfect, useless group. Choosing a sensible number is a genuine decision.
Real-world applications
Regression is everywhere a number must be estimated: forecasting sales and demand, pricing houses and ad inventory, predicting delivery times, estimating a patient's risk score, projecting energy load. If the deliverable is "how much" or "how many," it is regression.
Classification drives most automated decisions: spam filtering, fraud flags, disease diagnosis from labeled scans, sentiment (positive / negative), credit approval, image and speech labeling. If the deliverable is "which one" or "yes / no," it is classification.
Clustering powers discovery and exploration: customer and market segmentation, grouping documents or images by similarity, anomaly detection (points belonging to no normal group), and compressing data for visualization. If the deliverable is "what natural groups exist here," it is clustering.
Your turn
The challenge gives you several described scenarios. For each, name the task type — regression, classification, or clustering — using the output-based test from this page. This is exactly the judgment you make at the start of every project.
For each scenario, decide whether it is regression, classification, or clustering. Use the output-based test: a continuous number -> regression; a category from a fixed set -> classification; discovering groups from UNLABELED data -> clustering.
Fill in the dictionary tasks so each scenario key maps to exactly one of
the strings "regression", "classification", or "clustering":
"predict_tomorrows_temperature"— predict tomorrow's high temperature in degrees, from past weather."is_email_spam"— label each incoming email as spam or not spam."segment_shoppers"— group shoppers into natural segments when you have NO predefined segments, only their behavior."predict_house_price"— predict a house's sale price in dollars."identify_flower_species"— predict which of three species a flower is, from its measurements."group_songs_by_audio"— organize an unlabeled music library into groups of similar-sounding songs.
The hidden tests check each individual answer.
Check your understanding
What is the single most reliable way to tell regression, classification, and clustering apart?
By the size of the dataset
By which scikit-learn module the algorithm lives in
By the output: a continuous number means regression, a category from a fixed set means classification, and discovered groups from unlabeled data means clustering
By the number of features in X
The iris target is stored as the integers 0, 1, and 2 for the three species. Is predicting it regression or classification, and why?
Regression, because the target values are numbers
Regression, because the model outputs values between 0 and 2
Classification, because those integers are just names for three distinct categories — there is no meaningful "species 1.5"
Clustering, because there are multiple species
Which task is unsupervised, requiring no labels?
Regression
Classification
Clustering — it discovers groups from the features alone, with no target y
All three are unsupervised
A bank wants to predict, for each loan applicant, the probability-driven decision of "approve" or "deny." What task type is this?
Regression, because probabilities are numbers
Clustering, because applicants form groups
Classification, because the output is a category (approve or deny) chosen from a fixed set
It is not a machine learning task
Why can't you compute "accuracy" on a clustering result the way you do for a classifier?
Because clustering is always perfectly accurate
Because accuracy is only defined for regression
Because clustering has no ground-truth labels to compare against — the group ids are invented by the algorithm, so there is no "correct" answer to score
Because clusters change every run, so accuracy is meaningless by chance
Which statement about choosing a task type is correct?
Each dataset supports exactly one task type, fixed by the data
The inputs (features) determine whether it is regression or classification
A single dataset can support several tasks — the task is set by the question you ask (predict a number, predict a class, or find groups), not by the data alone
Clustering and regression are interchangeable for any dataset
Supervised 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.
The scikit-learn API
scikit-learn's quiet superpower is consistency. Every model — linear regression, nearest neighbors, k-means, and hundreds more — wears the same interface. Learn fit, predict, transform, score, and predict_proba once, and you know how to drive them all.