A Graphical Method for Assessing the Results of Fuzzy Clustering

• Activity Number: 339
• Type: Contributed
• Date/Time: Wednesday, August 6, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Graphics
• Author(s): Patrick Bobbitt*+ and Sumer J. Yates
• Companies: Bureau of Labor Statistics and Bureau of Labor Statistics
• Keywords: clustering ; fuzzy clustering ; K-Means clustering

Abstract

When attempting to ascertain the clustering of groups in the absence of other information, the standard technique has been to use "typical" clustering methods such as K-Means clustering to assess the closeness; then on the basis of an arbitrary criterion of "close," assign elements that are close to the same grouping. There are some applications in which this method could be seen as taking data reduction too far because it reduces all of the information of closeness into a series of indicator variables, and gives no information to the end user as to possible "leanings" into "neighboring" groups. To address this issue, a method of fuzzy clustering was developed that assigned to each data element a probability of inclusion into an arbitrary number of groupings (effectively a weighting that is applied to the indicator variables). The output from this method, unfortunately, leads to more data that may be difficult to quickly interpret. We have developed a method whereby the probabilities of inclusion are mapped to a set of primary colors whose mixture will allow the end-user a quick summary of the information produced by the fuzzy clustering algorithm.