When Two Datasets Disagree

The spreadsheet opens. The columns join. The chart looks convincing.

Then someone asks whether both datasets measured the same thing.

That question belongs near the beginning of data work, although it often arrives after the presentation has been made. Library practice offers a useful habit here: identify the subject carefully before bringing descriptions together. Shared spelling is weak evidence of shared meaning.

Scientific ontologies developed under the pressure of this problem. Researchers needed to compare findings produced by different communities without requiring every community to use identical local language.

Biology needed common descriptions

The Gene Ontology Consortium’s 2000 paper brought together researchers including Michael Ashburner, Judith Blake, and Suzanna Lewis. The project developed shared vocabularies for describing gene products across organisms, organized around molecular function, biological process, and cellular component. Those are different aspects: what a gene product does, the larger activity it participates in, and where it acts. Gene Ontology Consortium, Gene Ontology: Tool for the Unification of Biology

That separation should sound familiar after Ranganathan. Several dimensions can describe one subject without being competing answers to one question.

The point was not to rename every organism or replace all the specialist databases. Common descriptions could connect research that would otherwise remain difficult to compare.

There is a recognizable professional parallel. A facilities team records equipment by its function, location, and component type. Those dimensions must remain distinct if the organization wants to ask which equipment performs a function in a particular environment. “Basement” cannot become a kind of pump just because both appeared in the same classification column.

Scientific ontologies make the relationships more explicit and formally usable. The older discipline of separating unlike kinds of description remains visible underneath.

A concept is not a finding

An ontology can define a biological function. Associating a particular gene product with that function makes a further claim.

The Gene Ontology distinguishes its ontology from annotations that connect gene products to terms. An annotation carries supporting information, including an evidence code. Experimental evidence, computational analysis, and other forms of support are distinguished rather than presented as interchangeable observations. Gene Ontology, Guide to Evidence Codes

That is a valuable distinction far beyond biology.

Imagine an organization defines the category “corrosion-resistant component.” The definition says what the category means. A supplier’s declaration that a particular product belongs in it is an assertion. A laboratory test may support that assertion under stated conditions. A language model extracting the supplier’s sentence adds another activity to the history; it does not add another physical test.

If the system collapses definition, claim, and evidence into one label, users lose the ability to examine why a product was classified that way.

An ontologist can preserve the layers. Keep the concept definition. Record who made the product claim, what evidence supports it, and the relevant conditions. Let a user distinguish a tested result from an inferred or unverified classification.

This is Briet’s documentary problem again. A record becomes useful evidence through its relationship to an object, an observation, and a question. Copying it into a second system does not create a second observation.

Making data reusable takes more than access

In 2016, Mark Wilkinson, Michel Dumontier, and their co-authors published the FAIR principles: Findable, Accessible, Interoperable, and Reusable. Their paper gave particular attention to making data and metadata usable by machines, including through identifiers, shared vocabularies, and qualified references. Wilkinson and colleagues, The FAIR Guiding Principles

A downloadable file is a useful beginning. Reuse also requires enough meaning to interpret it.

Suppose one dataset records a project’s cost at approval and another records its final outturn. Both columns are called “cost.” Converting the currency does not make the values equivalent. The event at which the amount was measured is part of its meaning.

Or consider a temperature dataset. A value needs a unit, but it may also need to identify the measured object, sensor location, time, and method. Room air temperature and equipment surface temperature can share a unit while answering different questions.

The ontology practitioner helps establish those distinctions and connect them to the data fields. Some mappings will be exact. Others will be broader, narrower, conditional, or impossible without additional information.

An honest non-match is useful. It prevents a tidy join from becoming a misleading comparison.

Definitions change, and analyses remember

A long-lived dataset cannot be interpreted entirely through today’s vocabulary. Categories can be split, merged, renamed, or retired as knowledge improves.

Consider an organization that divides a broad “water incident” category into leakage, flooding, and contamination. Historical records may not contain enough detail to assign every old incident to one of the new classes. Automatically distributing them would invent precision.

Keep the original classification and its version. Add a reviewed mapping where the evidence permits one. Preserve an unresolved category where it does not. An analyst can then explain why an apparent trend partly reflects a changed classification practice.

This is the maintenance problem we met with library headings. The scale and application have changed, but the need to preserve meaning through revision has not.

The strongest through line in this history is a working habit: ask what a description refers to, how it was assigned, and what a reader can reasonably do with it. Data science benefits when those questions are answered before modelling begins.

Generative AI can now propose classifications and relationships at remarkable speed. That makes the distinction between a useful suggestion and an accepted claim more important. The final part of this series returns to the library principles with that new participant in the room.