A polyhierarchy is a taxonomy in which a concept can have more than one broader parent. In a plain hierarchy every term has exactly one place. In a polyhierarchy a term can sit under two or more categories at once, and both placements are correct.
The standard illustration is a tomato. Botanically it is a fruit. In a kitchen it is a vegetable. Forcing one parent makes the taxonomy wrong for half the people using it. Letting it sit under both makes it right for everyone and slightly harder to maintain.
That trade is the whole subject. Polyhierarchy buys accuracy and costs you clean roll-up.
The test is whether a single parent would be a lie. Inconvenient does not qualify. Debatable does not qualify. It has to be wrong.
What the four have in common is that the second parent is imposed from outside. A regulator, a clinical standard or a market defines it, and the business cannot dissolve the overlap by deciding to. That is the signal worth watching for. When the overlap is internal and negotiable, somebody can usually just settle it, and a single parent becomes honest again.
If none of those shapes describe your domain, you probably want a facet instead. Independent attributes that happen to cut across your tree are a faceting problem wearing a polyhierarchy costume. Hierarchical vs faceted taxonomy covers that distinction.
Three things get harder, and the first one is the one that surprises people.
| Cost | What happens | Mitigation |
|---|---|---|
| Roll-up becomes ambiguous | A concept under two parents gets counted twice, or once, depending on the query | Name a primary parent for reporting, and record that it is a reporting convention |
| Navigation duplicates | Users reach the same concept by two paths and wonder whether they are the same thing | Show the other parents on the concept page so the overlap is visible |
| Cycles become possible | A chain of broader relationships can loop back on itself | Validate on write; most tools will not catch this for you |
The double-counting problem is where polyhierarchies actually fail in production. Two reports built on the same vocabulary return different totals, both internally consistent, and the reconciliation takes longer than the modelling did. The fix is not technical. Somebody decides which parent is primary for reporting, writes that down, and the decision is honoured by every query.
SKOS supports polyhierarchy directly. A concept can carry more than one skos:broader relationship, and the standard expects it.
What SKOS deliberately does not do is enforce a tree. It has no notion of a single correct parent, which means the structure you express is the structure you meant. That is a feature when the domain is genuinely overlapping and a hazard when nobody is checking, because an accidental second parent looks identical to a deliberate one.
SHACL is where the checking lives. A shape can require that every concept has at least one broader term, that no broader chain forms a cycle, and that concepts in certain branches carry exactly one parent. What is SKOS covers the vocabulary and what SHACL is covers the validation side.
Many taxonomy tools assume a single tree, and they do not always say so on the box.
The export question is the one to insist on. A polyhierarchy that cannot leave the platform with its relationships intact has become a migration project the first time anyone wants to change tools. Taxonomy management software covers what to test more broadly.
Take a chemicals manufacturer classifying ten thousand products. The sales taxonomy groups by application: coatings, adhesives, lubricants. The compliance taxonomy groups by hazard class, which is set externally and has nothing to do with application.
A single tree forces a choice between them, and each choice breaks somebody's job. Sales cannot find products by what they are for, or compliance cannot answer an audit. Facets would work if hazard class were a free attribute, and it is not: it carries its own internal hierarchy of categories and subcategories that regulators define.
So the honest structure has products sitting under both an application parent and a hazard parent, each of those parents being a tree in its own right. That is a polyhierarchy, and it is the right answer here because both parent structures are real, externally defined, and neither can be demoted to an attribute.
None of those are modelling questions. They are governance questions that the modelling makes unavoidable, which is the usual pattern with polyhierarchy.
Who decides when a second parent is allowed?Left ungoverned, polyhierarchy spreads. Every disagreement about where a concept belongs resolves as "put it in both", and within a year the structure has no shape at all.
The working rule is that a second parent is a decision, with the same review as any other change to meaning. Somebody proposes it, somebody with standing in that domain approves it, and the reason is recorded alongside the concept. A term with three parents and no recorded reason for any of them is not a polyhierarchy. It is an argument nobody finished.
Taxonomy governance covers how that review is usually structured and what the trail needs to hold.
Use a polyhierarchy when a single parent would misrepresent the domain, and not merely when placement is contested. Expect to pay for it in reporting, and pay that cost deliberately by naming a primary parent for roll-up.
TopQuadrant's view is that polyhierarchy is safe when it is governed and corrosive when it is not. The structure is expressible in SKOS and checkable in SHACL, which means the question of whether a second parent was agreed can be answered from the data instead of from memory. Build it that way. Then the overlap is a decision somebody owns, and not a place where the taxonomy quietly stopped having a shape. Taxonomy management covers the surrounding discipline.