Supervised vs. Unsupervised Learning
Two ways to learn
Every method for learning from data starts by answering one question: does anyone already know the right answer? If they do, and that answer is written down alongside each observation, the method can learn to reproduce it. That's supervised learning, and it's most of what this course covers. But sometimes nobody knows the right answer; there simply isn't one recorded anywhere. All that's left is the raw measurements themselves, and the task quietly changes from “predict the label” to something stranger: decide whether these observations sort themselves into groups at all, with no one checking your work. ISLP calls this working blind , the second kind of statistical learning there is .
These aren't the same task with a detail hidden; they're genuinely different questions. Supervised learning answers “is this one U or V,” a question with a correct answer built into it from the start. Unsupervised learning can only ever say “these seem to belong together, and those seem to belong together” — it has no way to know what U or V even are, or whether the grouping it stumbled onto means anything at all.
Bringing this to life
Picture a retailer sitting on a spreadsheet of two numbers per customer: how much they spend a month, how often they buy. Nothing else, no tags, no segments. Hand that same spreadsheet to two analysts with two different assignments, and ask each one whether a given customer counts as Premium or Standard, and you get back two entirely different kinds of answer . That split is exactly what the two panels below are built to show, live.
The customers behind both panels are identical, down to the coordinates. The left one already knows Premium from Standard for every point, and fits a logistic regression to predict it, gradient descent nudging the boundary into place one step at a time. The right one gets the same two numbers and nothing else, no segment, no hint that “Premium” and “Standard” are even categories, and has to decide for itself, via k-means, whether two groups are there at all. It never borrows the left panel's colors while it works, that would hand it information it doesn't have; only after you reveal the true segments does color carry any meaning on that side, and only because we're the ones checking its answer against a key it never saw.