Structuring Interactive Cluster Analysis
Speaker: Wayne Oldford (Dir. of Computational Math, UW)
The problem of cluster analysis, or finding groups in data, is inherently ill-posed;
hence the multitude of different methods which purport to solve ``the'' problem. In
this talk, a variety of examples illustrate this point and cast doubt on whether a
single universally useful clustering method exists.
Interactive statistical graphics are used to explore high dimensional data in search of
unanticipated structure in data. Some of these methods will also be illustrated.
The goal of the present research is integrate different clustering methods with
interactive data visualization so that cluster analysis can be more exploratory, mixing
and matching methods as they suit the observed data and the problem context.
The methodological approach taken here is to consider the interactive clustering
problem as one of looking at partitions of data as the primary objects of interest and
so to consider how one might navigate the space of all possible partitions.
Computational resources are dedicated toward making this navigation interactive and
intuitive. A prototype organization based on interactive graphics and navigating the
space of partitions will be illustrated.