Walk into a supermarket looking for something cold to drink. Soft drinks is a useful sign. It brings together cola, sparkling water, juice, and energy drinks.
But that is not the only way those bottles could be grouped.
A shopper might group them by flavour. A buyer by supplier. A nutritionist by sugar. A warehouse by the shelf they occupy. A recall team by batch. Each group is sensible. Each answers a different question.
The bottles did not choose one of those groups for themselves.
That is what I mean when I say a category is a decision.
A group needs a job
We make groups all day without thinking about it. Photographs go in albums. People become neighbours, colleagues, customers, or family. The same person belongs to all of those at once without becoming a different person. The group changes because the purpose changes.
This is easy to miss in a marketplace. Someone searching for Coca-Cola Classic may reasonably want cans, bottles, multipacks, and perhaps Coke Zero in one set of results. They are exploring.
Someone placing an order needs a narrower answer. A 330ml can is not a 500ml bottle. A multipack is not one can. The products may be related, but they are not interchangeable.
Someone handling a recall needs a different group again. The question is not which products look alike or sell together. It is which physical items came from the affected batch.
| The job | A useful group | What must not be assumed |
|---|---|---|
| Browse a marketplace | related products | that they are the same product |
| Buy an item | interchangeable offers | that every related product will do |
| Restock a shelf | products from one supplier or range | that customers see the same group |
| Run a recall | items from one batch | that the whole product family is affected |
There is no single, natural group called Coca-Cola products. There are useful groups for particular jobs.
A bookshop makes the same point without a barcode in sight. One novel can be shelved under crime fiction, displayed with its author’s other books, recommended to someone who liked something similar, and counted with its publisher’s stock. The book has not changed. What changes is the question being asked of it.
Similarity finds. A category explains.
Similarity is not transitive makes one part of this clear. Similarity can bring things close enough to consider together. It makes a candidate group.
A candidate group is a shortlist. It may contain the right item, its neighbours, useful alternatives, and a few false leads. Search results work this way. So do recommendations such as “because you watched this”. Their value is that they save us from starting with the entire catalogue.
A category makes a stronger claim. It gives a group a name and a reason to exist beyond one search. Soft drinks, film franchise, household, and product family can be stored, reused, changed, and discussed.
That means somebody has to be able to answer four questions:
- What is this group for?
- What can belong to it?
- Who decides when it changes?
- What happens to the things that no longer fit?
Without those answers, a category can still look neat in a spreadsheet. It just has no stable meaning.
When the category arrives as a tag
Some groups are designed by the marketplace. Books, crime fiction, children's books. Everyone uses them, which makes browsing and reporting easier.
But the useful language often comes from the person adding the item. A bookseller writes kids' mystery, books for curious readers, ages 8 to 12, or middle grade, and may know something the supplied menu did not anticipate.
Those tags are worth keeping. They are also not automatically shared categories. Kids' mystery may point towards children’s crime fiction. It does not settle an age range, a reading level, or where the book belongs in every bookshop. The same words get used differently by different sellers.
This is the same shape as letting someone type a free-text address rather than pick one from a dropdown. The words they enter are their description of a real place. A system can tidy them, compare them with known addresses, and suggest a likely match. It should keep the original words, and it should not pretend a plausible suggestion is a verified address.
A threshold can build the candidate group. It can say kids' mystery and children's mystery are close enough to inspect together. It cannot create a public category on its own, because that turns a guess about words into a decision that changes search, reporting, and what sellers are allowed to say.
The safer sequence is dull and it works. Keep the tag and where it came from. Suggest an existing category when the match is clear. Ask when it is not. When the same unmet idea keeps returning, define a new concept deliberately, using the four questions above.
The address book version
An address looks like a fact about a place. It is not. Somebody made it, for a purpose, and the purpose is still in there.
Deirdre Mask’s argument, which I have been keeping notes on, is that house numbers were not invented to help you find your way. They were invented to help the state find you. In eighteenth century England most men shared one of about eight first names, so a name identified nobody outside the village that already knew him. Rulers demanded permanent surnames. Then they numbered the houses. A peasant’s door was numbered the same way as an aristocrat’s, which was genuinely Enlightenment, rational and equal, and also made everyone taxable, conscriptable, and findable.
So: grouped for what? For legibility to the state. Post, deliveries and directions came later, layered on top of a decision made for a different job.
Manhattan makes the same point with a price on it. New York runs a vanity address programme. For around eleven thousand dollars a developer can apply to the borough president for a better address than the building’s door deserves, so a tower whose entrance opens onto a numbered side street is sold as a Park Avenue or Central Park address. The bricks do not move. The door does not move. The category moves, and a Park Avenue address is reckoned to add five to ten per cent to the flats inside.
That is this whole note in one transaction. If an address were a property of a building, you could not buy a different one. What is being bought is membership of a group, and the membership is worth money because other people use the group to decide things.
The household is where this gets closest to home. Two people share an address. That can make them one household for a delivery, an energy account, or a family photograph. It does not make their bank records, medical histories, or identities one thing.
Communities need more care still. People can be connected by where they live, what language they speak, when they arrived, their work, income, faith, family, or a shared need. These may be real patterns. None is a complete description of a person, and none tells you what one particular person wants, believes, or can afford.
A service might use a language group to decide which translations to offer. Good. It should not use that group to decide what one person understands. A charity might use an area to decide where to offer support. It should not conclude that every household there has the same need.
The category can help us notice where help may be useful. It must not become a substitute for seeing the person.
A category is not an identity
This is where the vocabulary from entities earns its keep.
An entity is the thing in the world. One can of Coca-Cola. One address. One person. An entity reference is a row, listing, or record that mentions it.
Entity resolution asks whether two references describe the same real-world thing. It is a narrow question, and a valuable one.
"Coke Can 330ML" ┐
"Coca-Cola Classic 330ml can" ├─ do these describe one sellable product?
"Coca Cola, 330 ml" ┘
If the answer is yes, the references can be linked to one product entity. Only then does it make sense to say that product belongs to a product family, a category, a supplier range, or a recall batch.
Those later links are relationships. They do not replace the identity decision.
| This is | It answers | It does not answer |
|---|---|---|
| A candidate group | What should I look at next? | What is the same thing? |
| Entity resolution | Do these records refer to one thing? | What should I browse together? |
| A category | What kind of thing is this for this purpose? | Who is this exact entity? |
| A relationship | How are these named things connected? | Whether the connection should define either one |
So a category is not an entity reference. It is a managed concept in its own right. A row saying Soft drinks is a reference to that concept. A row saying Coke Can 330ML is a reference to a product. The two can be linked without being confused.
The boundary is part of the data
We treat categories as if they were found rather than made. A database has a category column, so the category starts to feel like a fact.
Sometimes it is close to one. A recall batch has a documented production boundary. A council ward has a legal one. Even then the group is defined by a rule, maintained by somebody, and capable of changing.
Others are much looser. Affordable, local, healthy, and similar to you need somebody to say what the words mean here. A useful system keeps that decision visible: the rule, the source, and when it last changed.
That is not bureaucracy for its own sake. It is how a person can question the group that is affecting them. You cannot argue with a boundary nobody admits to drawing.
The sorites paradox is the reminder. Take one grain from a heap and it still looks like a heap. Keep going and the word stops fitting, though no single grain announced the change. Categories have that same soft edge. A system should be able to say so.
Was it ever in the data
Everyone has met the form where none of the options is you. Job title, relationship, sector, company size. You pick the nearest one and move on.
Two things happened there and only one was recorded. You made a choice, and you rounded. The system stores the choice at full confidence and keeps no trace of the rounding, so your record is now indistinguishable from someone the box actually fitted.
That is worth asking of any category already in use, and it splits into three questions.
Where did this group come from? There are three honest answers. It was observed, because somebody looked at how things actually cluster. It was decided, because a report or a team structure needed it. Or it was inherited, from a vendor, a standard, a regulator, or the system before this one. All three can be legitimate. The trouble is that a category column records none of them, so a group that came out of a workshop looks exactly like one that came out of the data.
Is it still doing the job it was made for? Categories are cheap to create and almost impossible to retire. Retiring one breaks reports, dashboards, saved searches, and every comparison with last year. So they accumulate, and the schema slowly stops describing the world and starts recording the order in which people asked for things.
Who does it not fit? Other is usually the most informative value in a column and the least examined. If one in ten records lands there, that is not noise. It is the shape of a category nobody has made yet. A category that is right for ninety per cent of records is not ninety per cent right. It is wrong about a tenth of the people it describes, and it is wrong about them quietly.
The procedure is already above, in the part about seller tags. Keep what the person actually said. Suggest a match. Ask when it is not clear. When the same unmet idea keeps returning, define a new concept deliberately. That is what a good marketplace does for a bookseller. It is much rarer for a person.
The point is not to stop grouping
We need categories. A marketplace without them is a warehouse with no signs. An address book without them is a list. A public service with no way of seeing patterns will miss the people who need it.
The point is to ask the question before the group starts doing work: grouped for what?
Similarity can find the possibilities. Entity resolution can keep references to one thing together. A category can make a useful, reusable distinction. None of them relieves us of choosing the purpose, drawing the boundary, and being honest about what the group cannot say.
Reading
- Deirdre Mask, The Address Book (2020), and my reading notes on spatial identity, where the house-number argument above comes from.
- Vanity address applications, New York City. An unusually honest document, in that it admits the address is a thing you apply for.
Related
- A place is not a point, on why a coordinate is not an identity either.
- Similarity is not transitive, on how a candidate group goes wrong when it is treated as a category.
- What kind of thing are you resolving?, for the entity and reference vocabulary.
- The sorites paradox.