Is this always supervised machine learning?

No. A model can be specified directly by decision makers, although some methods learn parameters from assignment examples.

Must every alternative receive a different position?

No. Many alternatives can share one category; category assignment is not a total ranking.

Can a high average compensate for every weakness?

Only if the selected policy permits it. Gates, vetoes, or other noncompensatory structures can prevent that.

AI Glossary / Mathematics and optimization / Intermediate

Multicriteria classification

Definition

Multicriteria classification is the assignment of alternatives to predefined categories using several evaluation criteria. In multiple-criteria decision analysis, the term commonly refers to sorting into ordered categories, such as insufficient, adequate, and strong. It expresses an evaluation policy across multiple attributes rather than merely predicting an arbitrary label from observed features.

What it means

A business may need to decide whether each proposed AI project is ready for a pilot. Ranking projects from first to last does not answer that question: even the highest-ranked proposal might lack essential preparation. Classification compares each proposal with category standards instead.
Ordered categories have a meaningful preference direction. Strong is above adequate, which is above insufficient. This differs from nominal labels such as translation, summarization, and search, which have no inherent better-to-worse order. The vocabulary varies across fields, so the decision-analysis meaning should be stated when discussing this term with machine-learning specialists.

How it works

Define the categories and the criteria used to describe each alternative. Specify whether larger or smaller criterion values are preferable, and establish how evidence is measured. Then choose a sorting model and the parameters that connect criterion performances to category assignments.
Possible models include an aggregate value with category thresholds, decision rules, or comparisons with reference profiles. Some allow good performance on one criterion to compensate for weakness on another; others limit compensation or include veto conditions. Parameters may be elicited from decision makers or inferred from assignment examples. Those examples encode a policy and can contain inconsistencies that need examination.

Three practical examples

A data team can assign datasets to readiness categories using completeness, documentation, and suitability for a particular task.
An operations team can group maintenance proposals into ordered preparation levels using evidence quality and implementation readiness.
A procurement team can classify candidate services against defined technical categories before deciding which eligible options to compare further.

Think of it this way

Think of placing work into labeled trays using a written checklist. Several items can belong in the same tray, and a new item need not push existing items elsewhere. The analogy is limited: sophisticated sorting models can represent uncertain comparisons and interacting judgments that a simple checklist cannot capture.

Worked example

Consider an original policy for classifying fictional training datasets. Each receives three scores between 0 and 100: completeness c, documentation d, and task relevance r. Higher scores are preferable. Define S=0.5c+0.3d+0.2r. Assign strong when S≥80 and d≥60; adequate when the strong rule fails but S≥60 and d≥40; otherwise assign needs improvement.
Dataset A has scores (90,50,90). Its aggregate is 45+15+18=78, and its documentation score is at least 40. It is adequate. Dataset B has (80,70,90), giving 40+21+18=79. It is also adequate, despite having a slightly higher aggregate score.
Dataset C has (100,30,100), giving 50+9+20=79. It receives needs improvement because documentation is below 40. Dataset D has (84,70,90), giving 42+21+18=81, and qualifies as strong. C and B have the same aggregate but different categories because the policy restricts compensation for weak documentation.
For a sensitivity check, improve A's documentation from 50 to 60 while holding its other scores fixed. Its aggregate becomes 45+18+18=81, so it moves to strong. By contrast, raising C's completeness cannot help because it is already 100; its documentation requirement must be addressed.
The threshold equalities are intentional: documentation exactly 60 satisfies the strong documentation gate. This is an illustrative weighted-score-plus-gates policy, not a claim to reproduce ELECTRE TRI. Its numerical weights and cutoffs would need justification before any real use.

Illustrative case study

In a fictional example, Willow Learning has eight candidate datasets for a language exercise. A relative ranking places one dataset first, but staff discover that its documentation does not support the intended use. They introduce explicit readiness categories and record which criterion prevents each dataset from progressing.
Several datasets enter the same category, and the team treats that as acceptable: the task is to identify preparation needs, not force a unique league table. This invented case demonstrates the distinction between category assignment and ranking; it does not establish that the illustrative policy is suitable for real datasets.

Common misconceptions and mistakes

Do not confuse classification with sorting a list by one numerical column. In this context, sorting means assigning alternatives to categories according to a multi-criteria model.
Do not assume that fitting historical assignments proves the policy is desirable. Past decisions can be inconsistent, poorly documented, or inappropriate for the new use. A fitted model may reproduce those weaknesses.
Category names are not enough to define a model. Boundary values, tie treatment, missing-data handling, and allowed compensation all affect assignments. If a model is intended to respect improvement on every criterion, check that an otherwise identical but better-performing alternative cannot receive a worse category.

Business use

Multicriteria classification helps organizations apply a consistent evaluation policy across many items and explain what improvement is needed. The output can support a workflow: different categories can trigger different review steps or preparation activities.
An effective report includes the category meaning, measured performances, relevant thresholds, and the reason for assignment. For borderline cases, show which uncertain measurements could change the result. Keep the policy version with the decision record, because changes in criteria or thresholds can change categories even when an alternative's underlying performance remains the same.

Check your understanding

In the worked policy, a dataset scores 82 overall but has documentation 35. Is it strong?

Answer: No. It fails both documentation gates and is assigned needs improvement. The aggregate score alone is insufficient to determine the category.

Related terms

Preference elicitation
Monotonicity
multiple-criteria decision analysis
ordinal classification
sorting model
ELECTRE TRI
reference profile
veto criterion

Learn next

Write a small decision table for the worked policy, including documentation scores exactly 40 and 60. Add one case with missing documentation evidence and decide how it should be reviewed before calculating a category. Compare the explanation produced by this policy with a simple overall-score ranking.

Sources

Primary research source: Denis Bouyssou and Thierry Marchant, An axiomatic approach to noncompensatory sorting methods in MCDM II: More than two categories: https://www.lamsade.dauphine.fr/CahiersLAMSADE/CahiersLamsade230.pdf . The introduction and framework for ordered categories were consulted. The dataset policy and cases above are original illustrations.

Previous: Multicollinearity Next: Multidimensional analysis