Every method in this course answers one of four kinds of question. That's true no matter how different it looks on the page. Naming those four kinds now helps later. Each later chapter's opening paragraph will land right away. It won't feel like a brand new, unrelated topic each time.
Ask yourself what kind of question you're actually asking about a new data point. "How much will this cost?" is regression. The answer is a number. "Which category is this?" is classification. The answer is a label, picked from a list. "Which of these things belong together?" is clustering. Nobody told you the groups in advance. The method has to find them. "Can I describe this more simply, without losing what matters?" is dimensionality reduction. That means squeezing a big pile of measurements down to the handful that actually carry the story. Every chapter in this course is really just a tool. Each one answers one of those four questions, dressed up in its own name and math.
Regression predicts a number: a house's price, tomorrow's temperature. Classification predicts a category: spam or not-spam, or which of five species. Clustering groups similar samples together. It uses no target vector \mathbf{y} at all. No "correct" grouping is ever given, only the features. Dimensionality reduction squeezes down a wide design matrix, one with many features, into a smaller one. That smaller one keeps most of the useful information. Again, no target vector is involved. Regression and classification are supervised. Every training sample comes with a known y. Clustering and dimensionality reduction are unsupervised. Only \mathbf{X} is given, and the method has to find structure on its own. This course's own chapters run supervised methods first: Chapters 6-11 cover regression, classification, trees, ensembles, SVMs, and testing. Then come unsupervised methods, in Chapter 12. Last are several applied topics that don't fit neatly into either camp, in Chapters 13-18.
Which of the four questions is this?
Sort each real-world scenario by the kind of ML problem it is.
Loading interactive…
Worked example
Classifying a task by what's given, not by its name
-
1
A streaming service wants to group its shows into "similar to each other" clusters, for a picks page. It uses only each show's genre tags and runtime. No human ever labeled which shows belong in which group. Is this supervised or unsupervised? Which of the four problem types is it?
-
2
No target vector exists. Nobody supplied a "correct" grouping to learn from. Only feature data — genre tags, runtime — exists about each show.
-
3\boxed{\text{Unsupervised, clustering}}
The absence of a \mathbf{y} is the entire test. A similar-looking task — "predict which genre a show belongs to," where genre IS the target — would instead be supervised classification.
Loading interactive…
Worked example
Full depth or a light pointer?
-
1
Which of these does this course teach in full? Which gets only a light bridge before pointing to a later course?
-
2
Supervised methods, like trees and combined models, are part of this course's real core. So are unsupervised methods, like PCA and clustering. Transformer models, and running trained models in the real world, are different. Both belong to courses named above.
-
3\boxed{\text{Full: random forests/boosting, PCA/clustering. Light: transformers, deployment infrastructure.}}
Loading interactive…
Where this shows up
This course teaches reinforcement learning in one paragraph only.
That's a different kind of problem again. A system learns there by
receiving rewards for actions, rather than from a fixed dataset. This
course names reinforcement-learning as its real home. Chapter 16 is a
light bridge into neural networks. It names deep-learning as where
that material is taught in full. Chapter 17 does the same for
deployment, naming ai-systems-mlops. Chapter 15 does the same for
applied text tasks, naming natural-language-processing. Every one of
those is a genuine, working method. It's just not this course's job to
teach past an honest introduction.
Check your understanding (counts toward your progress)
A colleague says "clustering is just classification without labels." Explain what's right and what's misleading about that description.
AI grading isn't configured on this deployment.
Right on try — credit. Right — the answer had already been shown, so no credit this time. Next try is worth Answer shown after three tries — this one no longer earns credit. You chose to see the answer — this one earns no credit.
First mistake
Correct answer
Also accepted:
Why
Why
The description captures a surface similarity but misses the deeper one -- classification learns to reproduce a known labeling scheme; clustering has no labeling scheme to reproduce, only a similarity structure to discover.
Which of these does this course teach in full depth, rather than as a light bridge or one-paragraph mention?
Right on try — credit. Right — the answer had already been shown, so no credit this time. Next try is worth Answer shown after three tries — this one no longer earns credit. You chose to see the answer — this one earns no credit.
Transformer mechanics belong to deep-learning; this course's neural-network chapter is a light bridge only.
Full deployment depth belongs to ai-systems-mlops; this course's own deployment chapter is intentionally light.
Reinforcement learning gets one paragraph here; its real home is the reinforcement-learning course.
Why
This course's own core is classical supervised and unsupervised learning -- everything past that (deep NNs, MLOps, RL, applied NLP tasks) is named and pointed onward, not taught in full.
Predicting a customer's exact spend next month (a dollar amount) is an example of
Right on try — credit. Right — the answer had already been shown, so no credit this time. Next try is worth Answer shown after three tries — this one no longer earns credit. You chose to see the answer — this one earns no credit.
Classification predicts a category from a fixed list, not a continuous dollar amount.
Clustering has no target at all -- this task has a clear numeric target (spend), so it's supervised, not clustering.
Dimensionality reduction compresses features; it doesn't predict an outcome value.
Why
Predicting a number (not a category) with a known target for training is the definition of regression.