Multidimensional scaling (MDS) is a technique for turning abstract measures of similarity or difference between objects into a visual map. If you have data showing how alike or unlike a set of items are, MDS places each item as a point in space so that similar items appear close together and dissimilar items appear far apart. The result is typically a two- or three-dimensional plot you can actually look at and interpret, even when the original data had dozens or hundreds of variables.
The core idea is intuitive: you start with a table of distances or dissimilarities between every pair of objects, and MDS finds coordinates for each object that reproduce those distances as faithfully as possible. Think of it like reconstructing a map of cities from nothing but a chart of driving distances between them. You don’t need to know the latitude and longitude in advance. MDS figures that out.
How MDS Works
MDS begins with a symmetric matrix of dissimilarities. Each cell in this matrix represents how different two objects are from each other. These dissimilarities might come from direct measurements (physical distances, response times) or from subjective ratings (“On a scale of 1 to 9, how similar are these two things?”).
The algorithm then tries to place all objects as points in a lower-dimensional space, usually two or three dimensions, so that the distances between points match the original dissimilarities as closely as possible. It does this by minimizing a measure of error called “stress.” In its simplest form, stress is the sum of squared differences between the distances in the map and the input dissimilarities. A stress value near zero means the map faithfully represents the original data. Higher stress means the low-dimensional picture is distorting some relationships to make everything fit.
Because perfectly reproducing all the original distances in just two or three dimensions is rarely possible, MDS is always a compromise. The algorithm finds the best approximation given the space it has to work with. Choosing more dimensions reduces stress but defeats the purpose of creating a simple, visual summary.
Metric vs. Non-Metric MDS
There are two main flavors of MDS, and the distinction matters depending on what kind of data you have.
Metric MDS (also called classical MDS) assumes the actual numerical values of your dissimilarities are meaningful. If object A and object B have a dissimilarity of 4, and objects C and D have a dissimilarity of 8, metric MDS treats C-D as exactly twice as far apart as A-B. It tries to preserve those exact proportions in the output map. This works well when your input data consists of real measurements like physical distances, genetic distances, or reaction times.
Non-metric MDS only cares about the rank order of dissimilarities, not their exact values. If C-D is more dissimilar than A-B, the algorithm ensures that C and D end up farther apart in the map than A and B, but it doesn’t worry about preserving the precise ratio. This makes non-metric MDS ideal for subjective data. When people rate how similar two things feel, the difference between a rating of 5 and a rating of 6 may not be the same as the difference between 1 and 2. Non-metric MDS sidesteps that problem entirely by working with rankings alone.
Checking the Quality of the Map
Once MDS produces a configuration, you need to know whether it’s any good. The primary tool for this is the Shepard diagram, a scatterplot that compares the original dissimilarities (on one axis) to the distances in the MDS map (on the other axis).
In a well-fitting solution, the points in a Shepard diagram fall along a smooth curve or a straight line. This means the map is doing a good job of translating the input data into spatial distances. If the points scatter widely or form an L-shaped or step-like pattern, the solution may be “degenerate,” meaning the algorithm has found a configuration that technically minimizes stress but doesn’t actually represent the data in a meaningful way. When you’re working with similarity data (where higher numbers mean more alike), the Shepard diagram shows a decreasing trend, since similar objects should be close together. For dissimilarity data, it shows an increasing trend.
Stress values also give a quick quality check. While there’s no universal cutoff, values below 0.1 are generally considered good, and values above 0.2 suggest the map is struggling to represent the data in the chosen number of dimensions.
What MDS Reveals in Practice
MDS has been used across psychology, marketing, neuroscience, political science, sociology, and many other fields. Its power lies in making hidden structure visible.
One classic example involves color perception. If you ask people to rate how similar pairs of color patches look, then feed those ratings into MDS, the resulting map closely resembles Newton’s color wheel. Blue lands near purple, far from yellow. The algorithm recovers the circular structure of color perception without being told anything about wavelengths of light.
In psychology, researchers have used MDS to explore how people categorize animals. When participants rate how similar different animals are to each other, the resulting map places dogs close to cats, with both far from wolves and tigers. The spatial layout reveals that “domesticity” functions as an implicit dimension in people’s mental categories, even though nobody mentioned it directly. Similarly, when applied to perceptions of crime, MDS maps tend to place violent crimes (murder, assault) on one end of a dimension and minor offenses (speeding, jaywalking) on the other, revealing a primary axis of perceived violence.
In marketing, MDS helps companies understand how consumers perceive brands relative to competitors. If customers rate the similarity of various brands, the resulting map shows which brands occupy similar mental territory and which stand apart. This can reveal positioning opportunities that survey questions alone might miss.
Music researchers have used MDS to map perceived relationships between composers. A study might ask listeners how alike Mozart sounds compared to Beethoven, Chopin, Vivaldi, and Bach, then use the resulting map to identify the perceptual dimensions driving those judgments (era, style, emotional tone).
Interpreting the Dimensions
MDS produces a map, but it doesn’t label the axes. Figuring out what each dimension represents is the analyst’s job, and it’s often the most interesting part. You look at which objects cluster together and which fall at opposite ends of each axis, then infer what property separates them.
In one study of door-knocker designs, MDS revealed two clear dimensions. The horizontal axis separated door-knockers with faces from those without. The vertical axis separated metallic colors (silver, gray) from warmer tones (gold, bronze). Neither dimension was specified in advance. They emerged from the similarity ratings themselves.
This interpretive step is both the strength and the limitation of MDS. It surfaces patterns that people may not consciously articulate, but the labels you assign to dimensions are always somewhat subjective. Two analysts looking at the same map might describe the same dimension differently.
How MDS Compares to Related Techniques
MDS belongs to a broader family of dimensionality reduction methods, each with slightly different goals.
Principal component analysis (PCA) also reduces complex data to fewer dimensions, but it works directly with the raw variables rather than with distances between objects. PCA finds the directions of greatest variance in the data, while MDS works from a distance matrix and focuses on preserving pairwise relationships. If your starting point is a table of dissimilarities rather than a spreadsheet of measurements, MDS is the more natural choice.
Sammon mapping is a variant that modifies the stress function to give extra weight to preserving small distances. It divides each squared error by the original distance, so distortions between nearby objects are penalized more heavily than distortions between distant ones. This makes Sammon mapping particularly useful for identifying clusters, since it prioritizes getting the local neighborhood structure right even if some long-range distances are slightly off.
More recent methods like t-SNE and UMAP have become popular for visualizing very high-dimensional data (such as gene expression data or neural network embeddings). These techniques share MDS’s goal of producing interpretable low-dimensional maps, but they use different mathematical approaches that can handle much larger datasets. MDS remains widely used because of its simplicity, interpretability, and the fact that it works directly from any kind of distance or similarity data you can define.