Is MDS the same as PCA?

No. Classical MDS on Euclidean distances has a close equivalence to PCA, but MDS also covers other inputs and formulations.

Does MDS require two dimensions?

No. Two dimensions are convenient for plotting; other dimensions may better represent the relationships.

Can I compare two maps by their x-coordinates?

Not without considering alignment and the meaning of the coordinates. Distances may be unchanged despite rotation or reflection.

AI Glossary / Deep learning / Intermediate

Multidimensional Scaling

Definition

Multidimensional scaling, or MDS, is a family of methods that represents objects as points so that distances between the points reflect supplied dissimilarities between the objects. It often produces a two-dimensional map for exploring relationships.

What it means

The starting information can be a table of pairwise differences rather than a conventional table of measurements. For example, people might rate how different pairs of product designs feel. MDS tries to arrange those designs spatially so that similar ones appear close together.

The map is a representation of the chosen dissimilarities, not a discovery of an object's true position. Its usefulness depends on what “different” means, how reliably it was measured and how much distortion the chosen number of dimensions introduces.

Despite its placement in this glossary's deep-learning category, MDS is a broader statistical method. It does not require a neural network, labeled outcomes or a model that predicts future observations.

How it works

An analyst first defines the objects and assembles a compatible dissimilarity matrix. A common matrix is symmetric, with zero dissimilarity on the diagonal. Missing comparisons and unusual values need treatment appropriate to the implementation.

Metric MDS seeks map distances that approximate numerical dissimilarities. Nonmetric MDS instead emphasizes their ordering, allowing a monotonic transformation of the input values. Stress-based algorithms adjust point positions to reduce discrepancies.

Classical MDS uses an eigenvalue-based construction from a centered matrix derived from squared dissimilarities. It has a different optimization formulation from ordinary stress minimization. The distinction matters when comparing software and reported fit.

After fitting, inspect distortion and sensitivity to settings. Rotation, reflection and translation can leave pairwise distances unchanged, so axis direction alone usually has no inherent meaning. A low-dimensional solution can conceal relationships that need more dimensions to represent.

Three practical examples

A design team maps perceived differences among packaging concepts to guide interviews about confusingly similar alternatives.
An analyst visualizes distances between documents, while checking whether the underlying text representation captures the intended topic differences.
A researcher compares maps from separate respondent groups to explore whether their similarity judgments differ.

Think of it this way

Think of reconstructing a map from a table of distances between places. The table constrains where the places can sit relative to one another, even if nobody tells you which direction is north.

The analogy has limits. Product perceptions or document differences may not behave like geographical distances, and a flat page may not fit them exactly. An attractive arrangement is therefore an approximation that needs a fit assessment, not a literal map of a hidden world.

Worked example

Checking distances instead of trusting the picture

Suppose three objects have dissimilarities AB = 2, BC = 3 and AC = 5. Positions A = 0, B = 2 and C = 5 on a line reproduce all three values exactly.

For a simple raw-stress calculation, sum squared discrepancies across the three distinct pairs. The result is (2 − 2)² + (3 − 3)² + (5 − 5)² = 0.

An alternative arrangement A = 0, B = 1 and C = 5 gives distances 1, 4 and 5. Its raw stress is (2 − 1)² + (3 − 4)² + (5 − 5)² = 2. Reversing the first arrangement to 0, −2 and −5 preserves its zero stress. This illustrates both fit and orientation ambiguity. Real datasets may have no exact representation in the requested dimension, and different normalized stress definitions should not be compared as though identical.

Illustrative case study

A fictional retailer, Meadow Goods, asks a sample of customers to compare six package designs. An MDS plot places two concepts close together, prompting managers to call them interchangeable.

The analyst checks the original pairwise ratings and discovers substantial disagreement among respondents. The average map hides a subgroup that sees those concepts as quite different. A follow-up review examines subgroup maps, fit and the reasons participants gave for their ratings.

The retailer uses the maps to organize qualitative research rather than to declare a winning design. It also avoids naming an axis “premium quality” merely because a few expensive products appear on its right-hand side.

Common misconceptions and mistakes

Every axis has an obvious real-world label. Coordinate directions often depend on an arbitrary orientation; interpretations need additional evidence.
Nearness proves membership in the same customer segment. MDS is not itself a clustering or segmentation validation procedure.
Nonmetric means no distance calculations are involved. It refers to how input dissimilarities are related to fitted distances.
A low stress value proves the original measurements are valid. It measures aspects of the representation, not the quality of the underlying construct.

Business use

MDS can help businesses discuss competitive positioning, product resemblance and patterns in complex collections. Its strongest role is often exploratory: it makes a set of comparisons easier to inspect and question.

A useful deliverable includes the definition of dissimilarity, sample or data source, algorithm variant, dimensionality and fit information. For repeated reports, check whether apparent movement reflects changed data, altered preprocessing or merely a rotated solution. Teams should also consider the growing storage and computation needed for pairwise comparisons as the number of objects increases.

Check your understanding

Answer: Two maps look reversed left to right but have exactly the same distances between every pair.

Which one fits the dissimilarities better?

Answer: Neither on that evidence. Reflection preserves distances, so a distance-based fit criterion gives the same result.

Related terms

Learn next

Create a three-object distance table and draw a matching arrangement. Then change one distance and examine whether the original line still works. Compare metric and nonmetric objectives before interpreting an MDS plot from a business dataset.

Sources

Scikit-learn, manifold learning and MDS formulations: https://scikit-learn.org/stable/modules/manifold.html Scikit-learn, MDS implementation reference: https://scikit-learn.org/stable/modules/generated/sklearn.manifold.MDS.html Mair, Groenen and de Leeuw, More on Multidimensional Scaling and Unfolding in R: smacof Version 2: https://www.jstatsoft.org/article/view/v102i10

Previous: Multidimensional analysis Next: Multieffector planning