Unsupervised Learning

The default way to teach a model is to hand it the answers. Take 10,000 email messages, have people mark each one spam or not, and train the model to reproduce those marks. Material with marks attached like that is training data.

Two costs come with this. One is human time: 10,000 items at ten seconds each is 28 hours. The other is harder. Some questions have no answer that can be written down in the first place.

Pagedeck hit that wall. Mia proposed sorting customers into a few groups by character, and Theo asked how many groups there were and what each one would be called. Nobody had decided. A departed customer can be marked as departed, but whether customers come in three kinds or five had never been defined by anyone. This approach exists for questions where the answer key can't be produced: hand over the data as it is and let the structure inside it come out.

Five steps.

1. Turn the question into a task shape. Is this clustering, grouping similar things together; anomaly detection, finding values unlike the usual ones; or reducing a set of attributes that has grown too large? "Let's understand our customers" can't be run. "Let's group accounts whose behavior is similar" can.
2. Choose the inputs. Describe each account as a few numbers. At Pagedeck that's weekly active days, documents created per month, average people a document is shared with, and comment count. What goes in largely decides what comes out, which makes this the step that matters most.
3. Put the values on the same scale. Document count runs from 0 to 2,000 while active days runs from 0 to 7. Fed in raw, document count dominates the result, and you have effectively grouped by that one attribute. Rescale every input to a common range before starting.
4. Run it and vary the number of groups. Look at three, five, and eight side by side. Nothing tells you the right count, so the question becomes at which count the groups separate most distinctly.
5. Have a person read them and name them. If group 3 logs in often but shares with an average of one person, that's the solo-notes group. Skip this step and the output goes nowhere, because a group number is not something anyone can say in a meeting.

Scoring works differently too. With no answer key there's no accuracy to compute. Instead you check two things: whether the groups genuinely differ, and whether other numbers split along the same lines once you've divided by group. Renewal rates of 92, 71, and 34 percent across three groups makes the division usable. All three in the low seventies means the split is tidy and worthless, and the honest move there is to change the inputs rather than to keep hunting for a story in the groups you already have.

Use it when the answer key can't be made, or when making it would cost too much. Nobody knowing how many customer groups exist is the first case; needing to sort 200,000 documents by topic when the topic list itself doesn't exist is the second.

For the expensive case there's a hybrid. Group the 200,000 first, have a person review a representative hundred per group and name them, then use those names as an answer key to train a model that classifies the rest. Human effort falls from 200,000 decisions to a few hundred, and the classifier that results can be scored on accuracy like any supervised model.

It's the wrong tool where the answer already exists. Whether this customer will cancel next month, whether this message is spam — past cases carry answers, and using them is far more accurate.

Three limits. First, interpretation is a human job and it wobbles. Change one input or the number of groups and an entirely different division appears, with no criterion for declaring either one correct. Without the input list and settings written down, you can't reproduce the same result three months later.

Second, it frequently repeats what you already knew. Run it and the groups turn out to split along paid versus free, or by account size. Not a wrong result, just not a new one. When that happens, remove the distinction you already knew from the inputs and run it again.

Third, personally identifiable information (PII) in the inputs creates exposure for nothing. Names and contact details rarely help the grouping and only raise the cost of a leak, so inputs stay limited to behavioral records and drop anything that identifies a person.

"This has to be unsupervised. None of us can produce the answer key."

The classification criteria don't exist yet, so nothing can be labeled.

"Five groups came out and they're just paid versus free. Drop plan from the inputs and rerun."

A distinction already known has dominated the result.

"Don't report group numbers. Come back with names."

Output that skipped the interpretation step can't be used.

Pagedeck grouped 3,200 accounts on four inputs. The first run tracked the pricing plan exactly and got thrown away; with plan and seat count removed, the rerun produced three groups. One used it alone with frequent logins, one had three or four people editing documents together, one produced documents in volume and almost never shared them.

Renewal rates across those three the following quarter came to 34, 92, and 61 percent. Sofia decided to start with the accounts in the first group. The groups themselves weren't an answer; deciding where to look first was the result.

The most common reading is that skipping labels makes the data work lighter. Labeling disappears, so the job looks smaller. In practice the work moves rather than vanishing: choosing inputs, rescaling them, and reading results to give them names fills the space. Left uncorrected, a team runs the tool with no preparation, receives groups that mirror the pricing plan, and concludes the method is useless.

Believing a run surfaces things you didn't know is the second misreading. "Finds structure on its own" sounds that way. But the groups that come out are hypotheses, not answers. Knowing a cluster of frequent solo users exists says nothing about why, or what to do. Judgment arrives only after calling ten customers in that group or checking it against renewal rates. What this method does is narrow down where to look first, which is worth a great deal when the alternative is looking at 3,200 accounts in no particular order.