Learning Without Labels

Marketing catches Iris after a planning meeting with a question that sounds harmless: "What kinds of customers do we even have?" Iris opens the customer table to look for the column that would answer it — and realizes there isn't one. Every problem in Chapter 4 came with an answer column: churned or stayed, fake or genuine, delivered in 38 minutes. This question has no such column, because nobody knows the kinds in advance. Finding out is the whole point of asking.

Learning useful things from data with no answers attached is called unsupervised learning. If Chapter 4's supervised learning was studying flashcards with the answers printed on the back, this is a stack of cards with blank backs — nothing to check yourself against, and yet the fronts can still be compared, sorted, and searched for patterns. This page pins down what those patterns are, why they're worth hunting for, and the honest catch that comes with having no answer key.

No Answers, a Different Goal

Supervised learning always chases a known target. Someone decided in advance what the answer column means — fraud, churn, minutes — and the model's one job is to predict it for new rows. Unsupervised learning has no target to chase, so it does something else entirely: it discovers structure. Instead of "predict this", the instruction is closer to "look at all of it and tell me what you find" — which groups exist, what patterns repeat, which rows are strange.

Picture sorting a stranger's enormous photo archive. Nobody handed you categories, yet piles form on their own as you work: beach photos, birthday parties, screenshots of receipts. The piles emerge from similarity — these photos look alike, those don't — not from any instruction. Unsupervised learning is that sorting, done by a computer, on rows of data instead of photos. The piles were never defined; they were found.

What Does "Structure" Mean?

Three kinds of findable things cover most of what data teams mean by the word. First, groups: rows that huddle together because their values are similar — customers who order the same way, at the same hours, for similar amounts. Second, oddities: rows unlike everything else in the table — one account behaving like no other account. Third, redundancy: features that move together so tightly they tell one story, not two — at Plateful, delivery distance and delivery fee rise and fall almost in lockstep.

Notice that none of the three needed a label. Similarity, strangeness, and features-moving-together are all visible in the raw rows themselves. The next two topics take the first two and run: clustering turns groups into customer segments, and anomaly detection turns oddities into a fraud alarm.

Why Bother Without Labels?

Because labels are the expensive part. Chapter 2 showed what an answer column really costs: someone has to know the truth for every row and write it down — thousands of times, correctly, and again when the world shifts. For some questions the label barely exists at all: nobody reliably knows which reviews are fake, and fraud confesses itself months late or never. Meanwhile the unlabeled rows pile up for free. Every order, click, and delivery lands in Plateful's tables labeled with nothing — and that is most of the data the company will ever have.

And sometimes the question itself has no label to attach. "What kinds of customers do we have?" isn't a prediction with a right answer waiting to be checked — it's an open question about what's in the table. For questions shaped like that, unsupervised learning isn't the fallback option. It's the only tool that even fits.

The Catch: No Answer Key

Here is the price, stated plainly. With no labels, there is no test-set exam either. Chapter 3's ritual — hold back some examples, grade the model against the known answers — needs known answers to grade against. When the data team finds five customer groups, no hidden column says whether five was right, or whether these are the true groups. There is no accuracy score to report, because accuracy compares predictions to truth, and nobody has the truth.

So unsupervised results are judged the only way they can be: by usefulness, and by humans sense-checking them. Do the groups make sense to people who know the customers? Do decisions based on them work better than decisions without them? That's a softer standard than a test-set score, and it is a genuine limitation — not sloppiness, just the honest shape of learning without answers. When the data team says "here's what we found" instead of "we're 93% accurate", this is why.