Cluster Analysis

"Cluster Analysis" is the generic term for multivariate methods which attempt to find structures ("clusters") in the data. These methods are mostly based on calculations of the distance of the observations in multidimensional data space. Basically, cluster analysis will give answers to one of the following three questions:
 

• How many classes (clusters) can be observed in a data set?
• Which objects belong to which classes?
• How consistent are the classes?

At right you see a plot of about 150 data of three different kinds of flowers (50 each) which clearly shows two clusters. Cluster analysis can help to find such clusters automatically.

The results of a cluster analysis are often displayed as dendrograms which show the multidimensional relationships as a two dimensional line plot. In general, cluster analysis methods can be grouped into several categories:
 

• hierarchical methods (agglomerative and divisive clustering)
  e.g.  dendrograms, minimal spanning trees
• non-hierarchical methods
  e.g. density estimators by potential functions, or exchange procedures such as k-means clustering
• display methods
  e.g. principal component plots, or non-linear mapping plots