Multidimensional scaling (MDS) is a statistical technique that takes information about how similar or different a set of items are and turns it into a visual map. Items that are more similar appear closer together on the map, and items that are more different appear farther apart. It’s a way of making complex relationships visible at a glance, reducing dozens or hundreds of data points down to a simple picture you can actually interpret.
The Core Idea Behind MDS
The easiest way to understand MDS is with a classic example. Imagine you have a table listing the driving distances between every pair of major cities in the United States. That table contains real spatial information, but it’s buried in rows and columns of numbers. MDS takes those distances and reconstructs a map where each city appears as a dot, positioned so that the distances between dots match the original driving distances as closely as possible. You’d end up with something that looks remarkably like an actual map of the U.S., generated purely from the distance data.
That’s an artificial example because we already know what the “map” should look like. The real power of MDS shows up when you don’t know the underlying structure. If you have data on how similar 20 different brands of cereal are to each other (based on consumer ratings, nutritional profiles, or taste tests), MDS can arrange those brands in a two-dimensional space and reveal clusters and patterns that weren’t obvious from the raw numbers. Maybe one group clusters around “healthy and expensive” while another clusters around “sweet and kid-friendly.” MDS surfaces that structure automatically.
What MDS Actually Does With Your Data
MDS starts with a matrix of distances or dissimilarities between every pair of items in your dataset. These don’t have to be physical distances. They can be ratings of how different two products taste, how genetically distinct two populations are, or how differently two survey respondents answered a questionnaire. Any measurement that captures “how far apart” two things are can serve as input.
The algorithm then assigns each item a position in a low-dimensional space, usually two or three dimensions so the result can be plotted on a screen. It adjusts those positions iteratively, trying to make the distances between points on the map match the original dissimilarity data as faithfully as possible. The output is a scatter plot where proximity equals similarity. Close dots are similar items. Distant dots are dissimilar items.
One important nuance: the axes on an MDS plot don’t have predefined labels. Unlike a standard chart where the x-axis might represent time and the y-axis might represent temperature, MDS dimensions don’t come with built-in meaning. It’s up to the analyst to look at how items are arranged and interpret what each dimension represents. If brands of soda spread out along one axis by sweetness and along another by price, those labels come from human interpretation, not from the algorithm itself.
Metric vs. Non-Metric MDS
There are two main flavors of MDS, and which one you use depends on the nature of your data.
Metric MDS (also called classical MDS) works with actual numerical distances. It tries to preserve the exact values between items. If item A and item B are twice as far apart as item C and item D in the original data, metric MDS tries to maintain that exact ratio in the map. This approach works well when you have precise, quantitative measurements of dissimilarity.
Non-metric MDS takes a more flexible approach. Instead of preserving exact distances, it preserves only the rank order. If item A is the most different from item B, and item C is somewhat different from item D, non-metric MDS makes sure that relationship holds in the map, but it doesn’t worry about the exact magnitudes. This makes it especially useful for data based on subjective judgments (like “rate how similar these two flavors are on a scale of 1 to 7”), where the precise numbers may not be reliable but the rankings are meaningful. Non-metric MDS uses an iterative process, starting with a distance matrix and repeatedly rearranging points until the rank order of distances in the plot matches the rank order in the original data.
How to Tell if the Map Is Accurate
Squeezing complex, high-dimensional relationships into a flat two-dimensional picture inevitably loses some information. The question is how much. MDS quantifies this loss with a measure called “stress,” which ranges from 0 to 1. Lower stress means a more faithful representation of the original data.
- Below 0.05: Excellent. The map is a near-perfect representation with virtually no risk of misinterpretation.
- Below 0.10: Good. The map reliably reflects the real relationships in your data.
- Above 0.20: Potentially misleading. Patterns in the map may not reflect actual patterns in the data.
- Above 0.35: Essentially random. The points have little relation to the original distances.
If stress is too high in two dimensions, adding a third dimension often helps. The tradeoff is that three-dimensional maps are harder to visualize and interpret, so analysts typically aim for the lowest number of dimensions that still produces acceptable stress.
How MDS Compares to PCA and t-SNE
MDS isn’t the only technique for reducing complex data to a simpler picture. Two common alternatives are principal component analysis (PCA) and t-SNE, and each one prioritizes something different.
PCA focuses on preserving variance. It finds the directions along which your data varies the most and projects everything onto those directions. It’s fast and works well when the most important patterns are captured by the biggest swings in the data. MDS, by contrast, focuses on preserving the distances between pairs of points. This makes MDS better suited when you care about how individual items relate to each other rather than the overall spread of the data. MDS also has the advantage of working directly from a dissimilarity matrix, meaning you don’t need the raw data, just the relationships between items.
t-SNE takes yet another approach: it prioritizes keeping nearby points close together. Items that are neighbors in the original high-dimensional space stay tight in the t-SNE plot, but the technique doesn’t worry much about maintaining accurate distances between items that are far apart. This makes t-SNE excellent for spotting clusters but less reliable for interpreting global structure. MDS, because it tries to preserve distances between all pairs (including distant ones), gives a more faithful picture of overall relationships.
Where MDS Gets Used in Practice
In marketing, MDS is the engine behind perceptual maps, which are visual tools that show how consumers perceive different brands relative to each other. A company might survey customers about how similar various car brands feel, feed that data into MDS, and get a map showing which brands compete directly and which occupy unique positions. The closer a brand sits to a customer segment’s ideal point on the map, the more likely those customers are to consider and choose it. These maps feed directly into segmentation and positioning strategies that guide advertising, pricing, and product development.
In genetics and bioinformatics, MDS helps researchers visualize population structure. By measuring genetic distances between individuals or groups, scientists can create maps showing how populations relate to each other. These MDS plots often mirror actual geography, with genetically similar populations appearing near each other, but they can also reveal unexpected relationships like historical migration patterns or gene flow between groups that are geographically distant. MDS is also used to explore gene expression data, helping researchers spot patterns across thousands of genes by reducing that complexity to a two-dimensional picture where clusters of similarly expressed genes become visible.
In psychology and cognitive science, MDS has a long history of mapping how people perceive and categorize things. Researchers can collect similarity ratings for a set of stimuli (colors, faces, emotions, musical tones) and use MDS to uncover the hidden dimensions along which people organize those perceptions. A classic example: when people rate the similarity of different colors, MDS recovers a circular arrangement that corresponds to the color wheel, revealing that our perception of color has a structure that can be captured in two dimensions.
In ecology, non-metric MDS is a standard tool for comparing biological communities. Researchers measure how different the species compositions of various sites are, then use MDS to plot those sites on a map. Sites with similar ecological communities cluster together, making it easy to see the effects of environmental gradients, pollution, or habitat type on biodiversity.