|
ClustanMDS ...
is our implementation of Bell Labs' MDSCAL and KYST, which we have translated and optimized for Windows - three clicks and you're configured.What is multidimensional scaling?
Multidimensional scaling is the problem of representing n objects geometrically by n points, such that the inter-point distances correspond in some sense to experimental
similarities or dissimilarities between the objects. The underlying hypothesis is that the dissimilarities and distances are monotonically related. A goodness-of-fit
criterion called "stress" measures the extent to which the rank order of the dissimilarities corresponds to the rank order of the distances, and the MDS
procedure computes a configuration of points that minimizes the "stress" function. The objective is to obtain a scatter diagram, or plot, of your cases in two or more dimensions which
describes the relationships between them and their clusters. For example, suppose you want to display graphically
the relationships between the 5-cluster model of Mammals according to the composition of their milk
. Simply highlight the 5-cluster model on your tree and run MDS. The cluster model will then be displayed on an MDS scatterplot (left).
In this example, each cross corresponds to a case (species), colour coded by cluster, and the 5 clusters are identified by their
exemplars (Orangutan, Zebra, Fox, Deer and Seal). The scatterplot also gives us the clue that there are two outliers (Elephant and Rabbit),
which show up at the 7-cluster partition as singletons. To run ClustanMDS, simply read in or compute a proximity matrix, then
...
1. Click "MDS" in the Prox menu ... 2. Click "Start" in the MDS dialogue ... 3. Click "Finish" in the MDS dialogue ...
... and your optimized MDS two-dimensional configuration is displayed. Hence our claim ...
three clicks and you're configured ! For further information about ClustanMDS, click below ... Applications
Features
Parameters Limitations Example
To order ClustanMDS with ClustanGraphics, click here
Clustan - A Class Act © 1998 Clustan Ltd.
|