What things are being classified?
Define the domain.
A taxonomy only makes sense relative to the population it is meant to organize.
Side 28
Classification
& Taxonomy
A study of how categories are built. Classification makes complexity manageable by drawing boundaries — and every boundary changes what becomes visible, comparable and actionable.
objects→features→boundaries→categories→consequences
06category forms
05boundary tests
06classification risks
28Side
Classification is a sequence of decisions.
Before objects can be sorted, someone must decide which features matter, how much difference is acceptable and what the categories are for.
Which properties matter?
Observed, measured or inferred?
Feature selection determines which similarities count and which are ignored.
What creates membership?
Threshold, prototype, lineage, function?
Different rules can classify the same objects differently.
Where does one category end?
Sharp or fuzzy?
Boundary cases reveal whether the classification reflects real discontinuity or imposed convenience.
What decision will use the label?
Describe, predict, regulate, search?
A classification useful for one task may be poor for another.
Categories can be built in fundamentally different ways.
Knowing the category form helps explain why some boundaries are strict and others remain negotiable.
Defined by necessary features.
Membership depends on satisfying specified conditions.
Defined by resemblance to a central example.
Some members are more typical than others without one decisive feature.
Overlapping similarities.
Members share networks of traits without all sharing one common essence.
Defined by what something does.
Objects with different structures may belong together because they serve the same role.
Defined by common origin.
Lineage-based systems classify by descent rather than surface similarity.
Defined for governance.
Categories may be designed for reporting, eligibility or regulation rather than natural structure.
Taxonomies organize categories into relationships.
A taxonomy is more than a list: it encodes hierarchy, nesting or another systematic relation among categories.
Hierarchy
Broad classes contain narrower classes.
Useful when categories naturally nest from general to specific.
Mutual exclusion
Can an item belong to more than one category?
Some systems require one label; others allow multiple memberships.
Exhaustiveness
Does every case have somewhere to go?
“Other” categories can preserve coverage while hiding structural weakness.
Granularity
How fine should distinctions become?
Too coarse loses useful difference; too fine makes the system difficult to use.
Versioning
How does the taxonomy change?
Category systems need revision rules when new cases appear or old distinctions become obsolete.
Sometimes categories are discovered from similarity patterns.
Clustering groups observations without preassigned labels, but the result still depends on representation, distance and algorithmic choices.
What counts as similar?
Euclidean, cosine and other distance measures emphasize different relationships.
Are features comparable?
Variables with larger numerical ranges can dominate clustering unless transformed or standardized.
Partition around centroids.
Works best when clusters are compact and roughly spherical under the chosen distance.
Build nested groupings.
Dendrograms reveal how clusters combine across different levels of granularity.
Find dense regions.
Density-based methods can detect irregular clusters and treat sparse observations as noise.
A cluster still needs meaning.
Mathematical separation does not automatically imply a substantively important category.
Boundary cases reveal the design of the system.
Cases that fit poorly are not merely annoying exceptions; they expose which assumptions are doing the classificatory work.
An object fits two categories equally well.
Ask whether multi-label membership is acceptable, whether the categories overlap conceptually, or whether the underlying feature space is continuous rather than discrete.
A new case fits none of the existing categories.
Do not automatically force it into “other.” It may indicate that the taxonomy is incomplete or that the domain itself has changed.
A threshold splits nearly identical cases.
Administrative thresholds may still be necessary, but interpret the classification as a decision rule rather than evidence of a natural discontinuity.
Experts disagree on the correct label.
Separate disagreement about facts from disagreement about category definition. Inter-rater reliability can reveal ambiguity without deciding which ontology is best.
Labels do things.
Classification changes search, resource allocation, identity, regulation and prediction. Once a category enters a system, people may adapt to it.
Categories make some patterns countable.
What cannot be classified may disappear from reports and dashboards.
Labels can control access.
Administrative categories can determine benefits, services, restrictions or review.
Labels become model inputs.
Classification errors can propagate into later scoring and decision systems.
People can adopt or resist labels.
Social categories may influence self-understanding and group boundaries.
Classification changes behavior.
Once actors know how they are categorized, they may optimize, evade or contest the rule.
Old categories persist.
Databases, laws and institutional routines can preserve classifications long after their rationale weakens.
Sorting Things OutBowker & Star · classification systems and consequences
Women, Fire, and Dangerous ThingsGeorge Lakoff · categorization and cognition
Categories and Conceptsconcept formation and representation
Pattern Recognition and Machine LearningChristopher Bishop · computational classification