Glossary » multi-dimensional scaling
MDS; a statistical technique useful for understanding the structure of a domain, especially useful in early stages of design. People are asked to rate pairs of concepts for their similarity, then these similarities are fed into the statistical algorithm to determine dimensions that effectively describe these patterns of similarities. Essentially, the idea is to determine a space in which each concept can be mapped so that similar items are close and dissimilar items are far away. The MDS algorithm seeks to find the fewest dimension in which such a space is possible.
As an example, consider an application for choosing baby names where you want people to be able to find an appropriate name as quickly as possible. You would ask potential users (expecting parents, for instance) to rate the similarities of large numbers of names, e.g. “On a scale of 1 (very similar) to 7 (very dissimilar), how would you rate the names John and Sarah?” After getting a large number of ratings, the MDS algorithm would then determine the dimensions that best describe the data. The algorithm won’t name those dimensions for you, but will simply specify a number of dimensions and where each name falls along each dimension. So for instance, you may find a dimension in which Joe, Bob, and Tom all fall at one end while Jenny, Amy, and Susan all fall at the other end. In this case, you would be likely to identify the dimension as gender. In the example of names, you might also find dimensions of nationality and pronunciation. Thus, your interface for selecting names may allow the users to select gender, nationality, and sounds to help them find an appropriate name, or you may allow the user to ask for “similar” names once they have found one they basically like.