Is the output a probability?

Not by definition. It is an aggregate score whose interpretation comes from the model and application.

Are the singleton capacities enough?

Not for an arbitrary capacity. Group values encode interactions and must also be specified or constrained.

Can an input improve without changing the result?

Yes. Max-min aggregation can have flat regions, as the worked example shows.

AI Glossary / Graphs and symbolic AI / Advanced

Sugeno integral

Definition

The Sugeno integral is an aggregation method that combines ordered input scores using minimum and maximum operations and a monotone set function called a capacity. The capacity assigns importance or sufficiency to groups of criteria, allowing the combination to reflect interactions between them. In its common normalized form, scores and capacity values share a scale from zero to one. It is especially relevant when ordering matters more than numerical differences.

What it means

Suppose several criteria describe an option, but averaging their ratings would suggest unjustified tradeoffs. A very high score on one criterion might not appropriately compensate for a weak combination elsewhere. A Sugeno integral offers a different way to combine the evidence.

Its capacity describes groups, not just individual weights. A pair of criteria can have an importance value that is not the sum of the two separate values. The output is an aggregate evaluation under that chosen capacity. It is not automatically a probability, a confidence estimate, or an objective measure of quality.

How it works

For a finite set of criteria N, define a capacity μ with μ(empty)=0 and μ(N)=1. Require monotonicity: if group A is contained in group B, then μ(A) cannot exceed μ(B). The capacity need not be additive.

Sort the input scores from smallest to largest. For each score, consider the group consisting of that criterion and all criteria with scores at least as high in the sorted ordering. Take the minimum of the score and that group's capacity. The Sugeno integral is the maximum of these candidate values.

This construction compares levels and group sufficiency using max and min rather than a weighted sum. Inputs must have a coherent common ordering and interpretation. Specifying the capacity is part of modeling: it may be elicited from preferences or learned from data, with constraints and validation appropriate to the application.

Three practical examples

Multi-criteria demonstrations: aggregate ordinal assessments of a service option while explicitly representing which combinations of criteria are sufficient for a favorable evaluation.
Interpretable classification research: transform features into local scores, aggregate them, and compare the result with a threshold, evaluating the learned model on held-out data.
Sensitivity analysis: compare candidate capacities to reveal whether a ranking depends on strong assumptions about interactions among assessment criteria.

Think of it this way

Think of checking whether some sufficiently important group of requirements all reaches a satisfactory level. A group's weakest score limits what that group can support. The analogy is bounded: the actual integral checks numerical or ordered levels and a defined capacity, not a committee's informal judgment that an option feels adequate.

Worked example

Use an original three-criterion example with scores A=0.2, B=0.7, and C=0.9. Define singleton capacities μ({A})=0.3, μ({B})=0.4, and μ({C})=0.5. Define pair capacities μ({A,B})=0.6, μ({A,C})=0.7, and μ({B,C})=0.8. The empty group has zero and the full group has one. These values are monotone because each larger group has capacity at least as great as its subsets.

The scores are already sorted. At level 0.2, the relevant group is {A,B,C}, so the candidate is min(0.2,1)=0.2. At level 0.7, the group is {B,C}, giving min(0.7,0.8)=0.7. At level 0.9, only {C} remains, giving min(0.9,0.5)=0.5. The maximum is 0.7, the Sugeno result.

The ordinary arithmetic mean is 0.6, showing that this is a different aggregation rule, not a disguised average. If A improves from 0.2 to 0.6 while B and C stay fixed, the candidates become 0.6, 0.7, and 0.5. The output remains 0.7: an improvement need not strictly increase the aggregate.

If instead B rises from 0.7 to 0.85, with A at 0.2 and C at 0.9, the middle candidate becomes min(0.85,0.8)=0.8. The result rises to 0.8. Every number here is illustrative; the capacity has not been estimated from real preferences.

Illustrative case study

Fictional case: a research team compares help-desk service packages using ordered ratings for coverage, responsiveness, and documentation. Stakeholders disagree about whether strong responsiveness alone should make a package acceptable.

The team constructs several candidate capacities and shows how the aggregate changes when responsiveness is considered alone or alongside coverage. It asks stakeholders to review the resulting comparisons before choosing a model. The fictional case illustrates preference elicitation and sensitivity analysis, not a validated purchasing recommendation or evidence that Sugeno aggregation outperforms other methods.

Common misconceptions and mistakes

The capacity is not necessarily a probability measure, so values for disjoint groups need not add. Do not insert arbitrary units such as seconds, currency, and star ratings without mapping them to a meaningful common evaluation scale.

The Sugeno integral differs from the Choquet integral and from Takagi-Sugeno fuzzy rule models. Sharing the name Sugeno does not make these the same calculation. Finally, monotonicity guarantees that improving inputs cannot lower the aggregate under a fixed valid capacity, but it does not require every improvement to change the result.

Business use

The method can help when a decision process needs explicit group-level preferences and restrained compensation between criteria. Its usefulness depends on whether stakeholders can explain and defend the scoring scale and capacity, rather than merely accepting a sophisticated formula.

A general capacity has values for many subsets, creating an elicitation and learning burden as the number of criteria grows. Consider whether a simpler constrained family expresses the intended preferences. Check output stability, held-out performance when learning, and whether plateaus in the aggregate suit the decisions being made.

Check your understanding

With scores 0.2, 0.7, 0.9 and sorted candidate values 0.2, 0.7, 0.5, what is the Sugeno result?

Answer: 0.7, because the final operation takes the maximum of the candidates.

Related terms

Ordinal scale
Monotonicity
Capacity
fuzzy measure
aggregation function
Choquet integral
multi-criteria decision analysis
max-min operation

Learn next

Recalculate the example after changing one pair capacity while maintaining monotonicity. Explain which group determines the result. Then compare the same inputs under an arithmetic mean and a Choquet integral to understand the different assumptions each method makes.

Monotonicity

Sources

Abbaszadeh and Hüllermeier, Machine Learning with the Sugeno Integral, finite definition and capacity-based aggregation: https://arxiv.org/pdf/2007.03046 Marichal, On Sugeno Integral as an Aggregation Function: https://ikojadin.perso.univ-pau.fr/kappalab/pub/MarFSS2000.pdf The numerical capacity and fictional case are original. Scope: normalised finite aggregation on a common ordered scale.

Previous: Sufficient statistic Next: Sum of squared differences