Must every input dissimilarity satisfy the triangle inequality?

Not necessarily for a stress-fitting procedure, but violations can prevent an exact Euclidean representation and affect interpretation.

Is it equivalent to principal component analysis?

Classical MDS on Euclidean distances has a close equivalence to PCA. That statement should not be transferred unqualified to every stress-based metric MDS fit.

Can two maps look different but be equally good?

Yes. Rotation, reflection and translation can preserve every fitted distance and therefore the stress.

AI Glossary / Probabilistic machine learning / Intermediate

Metric multidimensional scaling

Definition

Metric multidimensional scaling represents objects as points in a chosen-dimensional space so their distances approximate supplied numerical dissimilarities. A common formulation adjusts the coordinates to minimize a stress function measuring discrepancies between target and fitted distances.

What it means

Suppose a team knows how different pairs of products, documents or respondents are, but wants a map that makes those relationships easier to inspect. Metric MDS uses the numerical sizes of those differences, not only their ordering.

The resulting coordinates are constructed from pairwise relationships. An axis does not automatically represent a named feature such as quality, cost or preference. Rotation, translation and reflection preserve distances, so equally valid maps can have different orientations.

Terminology varies across the MDS literature. This entry focuses on stress-based metric MDS with fixed target dissimilarities. Classical MDS is closely related but uses an eigendecomposition and a different fitting criterion; nonmetric MDS principally preserves ordering through an admissible monotonic transformation.

How it works

Start with a symmetric table of dissimilarities and choose the output dimension. Place points at initial coordinates and compute their Euclidean distances. Compare those distances with the target dissimilarities.

A common raw stress objective sums weighted squared differences over distinct pairs. An iterative method such as SMACOF updates coordinates using a majorisation procedure that does not increase the objective when its update assumptions are met. This monotonic improvement does not by itself guarantee finding the best possible configuration from every starting point.

Inspect stress, pairwise residuals and sensitivity to initialization and dimension. Dissimilarities may be noisy or impossible to reproduce exactly in the requested space. Different stress normalizations and weighting conventions produce different numerical scales, so scores require their definitions alongside them.

Three practical examples

A product researcher maps pairwise customer dissimilarity ratings to explore which products are perceived as close alternatives.
A document analyst visualizes a distance matrix derived from consistent text representations, checking whether the two-dimensional map distorts important pairs.
A materials team represents samples using pairwise laboratory dissimilarities when no single original measurement gives a useful overview.

Think of it this way

Think of arranging towns on a tabletop using a list of distances between them. The task is to make the separations on the tabletop agree as closely as possible with the list.

The analogy has limits: the supplied values need not be geographic distances, and a low-dimensional arrangement may be unable to reproduce every relationship. The table's horizontal direction has no inherent meaning without additional interpretation.

Worked example

Three equal distances cannot fit exactly on a line

Consider three objects A, B and C with target dissimilarity 1 for every pair. In two dimensions, an equilateral triangle with side length 1 gives zero stress.

Now require a one-dimensional representation. Choose the ordering A, B, C and place the points at (−a, 0, a), where a is nonnegative. The fitted distances are a, a and 2a. With each pair counted once, raw stress is 2(a − 1) squared + (2a − 1) squared.

Differentiating gives 12a − 8, so the minimum occurs at a = 2/3. The fitted distances are 2/3, 2/3 and 4/3. Their squared discrepancies sum to 1/9 + 1/9 + 1/9 = 1/3.

By comparison, coordinates (−1, 0, 1) produce distances 1, 1 and 2, giving stress 1. Matching two pairs exactly is worse than distributing the unavoidable error across all three.

Any ordering is equivalent here because all target distances are equal. Translating or reflecting the optimal line configuration preserves its stress. The nonzero minimum reflects the dimension restriction, not necessarily a programming mistake.

Illustrative case study

A fictional software company collects pairwise dissimilarity ratings for several service packages. Its research team uses metric MDS to prepare a two-dimensional exploration map.

Before naming the axes, the team checks pairwise residuals and compares the map with a three-dimensional fit. It discovers that a pair appearing close on the flat map has a substantial distance error, so the presentation includes that limitation rather than interpreting the apparent closeness as a strong customer preference.

The company treats the map as a summary of the chosen ratings. It does not claim that the geometry establishes market segments, predicts purchasing or identifies causal product attributes.

Common misconceptions and mistakes

Metric MDS only cares about ranks: that describes the central aim of nonmetric MDS; metric fitting uses numerical dissimilarity information.
Lower stress proves the dissimilarity measure is meaningful: it measures representation fit, not the quality of the underlying measure.
Axis signs carry fixed meaning: reflections preserve the same distance relationships.
A converged iterative solution must be globally optimal: different starts can reach different configurations or local solutions.

Business use

Metric MDS can make complex pairwise information understandable, provided the display does not hide large distortions. It is useful for exploration, communication and checking patterns rather than automatically deciding clusters or rankings.

Document how dissimilarities were obtained, scaled and weighted. Ask whether their numerical intervals are meaningful enough for metric fitting. Compare dimensions and inspect which pairs are poorly represented. If decisions depend on a particular relationship, consult the original dissimilarity as well as its plotted distance. Large datasets also create storage and computation costs because pairwise tables grow quadratically.

Check your understanding

A map matches all target distances, then every coordinate is multiplied by two. Does raw metric stress necessarily remain zero?

Answer: No. The fitted distances double while the fixed target dissimilarities remain unchanged. Unlike translation or rotation, rescaling generally changes the raw stress in this formulation.

Related terms

Multidimensional Scaling
Euclidean distance
Dimensionality reduction
Principal component analysis
nonmetric MDS
classical MDS
stress
SMACOF
dissimilarity matrix

Learn next

Build a three-object dissimilarity table and compare line and triangle representations. Calculate pairwise residuals rather than only a total score. Then compare metric, nonmetric and classical MDS on the same small example, recording the objective each method optimizes.

Sources

Scikit-learn user guide, multidimensional scaling and distinctions among formulations: https://sklearn.org/stable/modules/manifold.html

Scikit-learn SMACOF documentation, stress majorisation and algorithm interface: https://scikit-learn.org/1.8/modules/generated/sklearn.manifold.smacof.html

The equilateral-distance calculation and fictional product study are original illustrations. Raw stress here counts each unordered pair once; other reported conventions may differ.

Previous: Metric Learning Next: Metric space