Clustering Objects on Subsets of Attributes

                      Jerome H. Friedman*
                      Stanford University

                     

                            ABSTRACT

A dissimilarity measure between objects, each characterized by the
same set of measured attributes (variables), is proposed for use
in cluster analysis. It assigns small dissimilarities to object pairs
that have close values on any subset of the attribute variables
regardless of their values on the complement set of variables. Using
this measure in conjunction with standard dissimilarity based
clustering algorithms encourages the detection of subgroups of
observations that preferentially cluster on subsets of the attributes,
without having to explicitly search for the relevant subsets. The
relevant variable subsets for each individual cluster can be different
and partially (or completely) overlap with those of other
clusters. Enhancements for increasing sensitivity for detecting
especially low cardinality groups clustering on a small subset of
variables are discussed. Applications to several different domains,
including gene expression arrays, are presented.

* Joint work with Jacqueline J. Meulman, Leiden University