Does the skyline always contain one option?

No. Several non-dominated alternatives can remain.

Does being in the skyline mean an option is best overall?

No. It means no supplied alternative dominates it on the selected dimensions.

Can identical preference values dominate each other?

Not under the usual strict-improvement requirement; duplicate handling is a separate issue.

AI Glossary / Data engineering and architecture / Advanced

Skyline operator

Definition

A skyline operator returns the records that are not dominated by another record across a chosen set of preference dimensions. A record dominates another when it is at least as good on every selected dimension and strictly better on at least one. The result preserves trade-off alternatives without requiring a single weighted score. It does not necessarily select one winner or rank the returned records.

What it means

Suppose a choice involves both cost and completion time. One option may cost less while another finishes sooner, leaving a real trade-off. A third option that costs more and finishes later than an available alternative has no advantage on those selected criteria. A skyline query removes such dominated options. Its conclusions are relative to the candidates, dimensions and preference directions supplied. If an important criterion is omitted, an excluded option might still matter to the actual decision. The operator simplifies comparison; it does not decide what the user should value.

How it works

The Skyline Operator paper by Börzsönyi, Kossmann and Stocker describes a database operation for retaining non-dominated tuples and discusses several implementation algorithms. Each chosen dimension has a preference direction, such as minimizing cost or maximizing quality. The dominance test combines comparisons across all those dimensions. A straightforward explanation compares each candidate with others, although efficient implementations can avoid many comparisons. The original paper proposes SQL extensions; that historical proposal should not be assumed to be syntax supported by every current database. The logical result is distinct from the particular algorithm used to compute it.

Three practical examples

An analyst filters hypothetical processing configurations using lower runtime and lower resource cost as the selected preferences.
A model-evaluation worksheet retains configurations for which no other measured configuration has both at least as high accuracy and no greater latency, with a strict improvement somewhere.
A planning tool presents non-dominated alternatives before users apply additional preferences or requirements to make a final choice.

Think of it this way

Imagine comparing backpacks by lower weight and greater capacity. If one backpack is no heavier and holds more than another, it has the advantage on those two measures. But a lighter backpack with less capacity and a heavier one with more capacity remain a trade-off. The skyline keeps such alternatives. It does not account for comfort or durability unless those factors are represented in the comparison.

Worked example

Consider four fictional job configurations, with both cost units and runtime minutes to be minimized. A has (10, 8), B has (12, 9), C has (8, 12), and D has (10, 7). A dominates B because 10 is less than 12 and 8 is less than 9. D dominates A because its cost is equal at 10 while its runtime is lower at 7. D also dominates B. C and D do not dominate one another: C costs less, but D finishes sooner. The skyline is therefore C and D, two of the four candidates. The calculation does not choose between them. If a separate requirement allows at most nine minutes, C becomes ineligible and D remains among these candidates.

Illustrative case study

In a fictional AI-service experiment, engineers compare measured accuracy and response latency for several model configurations. They use a skyline filter to identify configurations that offer distinct trade-offs on those two measures. During review, a colleague notes that memory use was not included and that one retained configuration exceeds the intended device's capacity. The team applies the capacity requirement and recomputes the eligible comparison. It also avoids presenting small accuracy differences as definitive without considering measurement uncertainty. The scenario illustrates how a skyline supports comparison while relying on appropriate inputs. It is not a benchmark or a claim that the returned configurations are universally best.

Common misconceptions and mistakes

A skyline is not the same as taking the top few rows after sorting by one column. It can return many candidates, especially when dimensions create many trade-offs. Nor does a retained point need to be best on any one dimension by itself. Another mistake is using the wrong preference direction, such as minimizing a quality measure that should be maximized. Equal values require attention too: strict improvement on at least one dimension is part of dominance, so identical preference values do not make either record strictly dominate the other. Handling duplicate records is a separate implementation choice.

Business use

The operator is useful when teams want to narrow a set of alternatives before agreeing on precise trade-off weights. Define eligibility requirements and comparison dimensions clearly, and explain what was excluded from the model. In AI evaluation, ensure that measurements were obtained under comparable conditions and that uncertainty is not hidden by a crisp dominance label. A skyline can make a review more manageable, but a final decision may need additional constraints, stakeholder preferences or evidence. Record those later decisions separately so readers can distinguish mathematical filtering from the judgment used to select an option.

Check your understanding

Which toy configurations remained?

Answer: C and D.

Why did D eliminate A?

Answer: It matched A's cost and improved its runtime.

Why did C remain despite taking longer?

Answer: Its lower cost prevented D from dominating it.

What changed when the nine-minute requirement was introduced?

Answer: C failed eligibility, leaving D among the stated alternatives.

Related terms

Multi-objective optimization
Pareto dominance
Pareto frontier
database query
preference query
decision support

Learn next

Create five fictional options with two criteria and state which direction is preferred for each. Test dominance pair by pair and identify the skyline. Then add an eligibility rule or a third criterion and repeat the exercise. Explain why the retained set changes without claiming that the operator has learned the user's complete preferences.

Sources

Stephan Börzsönyi, Donald Kossmann and Konrad Stocker, The Skyline Operator, ICDE 2001: https://islab.kaist.ac.kr/Lecture/wst500_2012_Spring/pdf/00914855.pdf The configuration calculation, backpack analogy and AI-service case are original educational illustrations. The paper supports the dominance definition and database operator; the entry does not present proposed historical SQL syntax as universally supported or equate non-domination with a single best decision.

Previous: Sky computing Next: Slack