On the Minimum Description Length (MDL) Principle for Hierarchical Classifications
Hierarchical clustering procedures such as single-, average-, or complete-link procedures produce a series of groupings of the data arranged in the form of a hierarchy, or tree structure. In most cases, the choice of where to “cut” the tree is left to the user. Occasional formal guidelines have usually been based on ideas of random sampling, but that assumption is often violated in the contexts in which cluster analysis is used. This paper explores the application of Rissanen’s MDL principle to derive possible guidelines for cutting the tree. These guidelines do not assume random sampling.
KeywordsComplete Linkage Minimum Description Length Aggregation Criterion Hierarchical Method Penalize Maximum Likelihood
Unable to display preview. Download preview PDF.
- Bryant, P. (1996): The Minimum Description Length Principle for Gaussian Regression. Working Paper 1996–08, University of Colorado at Denver, Graduate School of Business Administration. Denver, Colorado 80217–3364.Google Scholar
- Duda, R. O. and Hart, P.E. (1973): Pattern Classification and Scene Analysis. John Wiley & Sons, New York.Google Scholar
- Everitt, B. S. (1993): Cluster Analysis. Edward Arnold, London.Google Scholar
- Johnson, R. A. and Wiehern, D. W. (1988): Applied Multivariate Statistical Analysis, second edition, Prentice-Hall, Englewood Cliffs, N. J.Google Scholar
- Rissanen, J. (1996): Shannon-Wiener information and stochastic complexity, In: Proceedings, N. Wiener Centenary Congress, East Lansing, Michigan.Google Scholar