COLLOQUIUM
         Computer Science Department, Boston University

Speaker: Jennifer Dy
         Northeastern University

Title: Clustering High-Dimensional Data

Date:  November 3, 2004
Time:  3pm
Place: MCS 135 (for directions, see www.cs.bu.edu/colloquium)

Abstract:

Creating effective algorithms for unsupervised learning is important
because vast amounts of data preclude humans from manually labeling the
categories of each instance.  In addition, human labeling is expensive and
subjective.  Therefore, a majority of existing data is unsupervised
(unlabeled).  The goal of unsupervised learning or cluster analysis is to
group ``similar'' objects together. "Similarity" is typically defined by a
metric or a probability model.  These measures are highly dependent on the
features representing the data.  Many clustering algorithms assume that
relevant features have been determined by the domain experts.  But, not
all features are important.  Moreover, many clustering algorithms fail
when dealing with high-dimensions.  In this talk, I will present three
approaches for dealing with clustering in high-dimensional spaces:
1.   Feature subset selection for unsupervised learning,
2.   Feature selection and clustering in an interactive visualization
environment; and
3.   Hierarchical feature transformation and clustering.
This talk explores the feature subset selection for unsupervised learning
problem.  We investigate the problem through our algorithm called FSSEM
(Feature Subset Selection wrapped around Expectation-Maximization
clustering) and through two different performance criteria for evaluating
candidate feature subsets:  maximum likelihood and scatter separability.
We identify two issues:  the need for selecting the number of clusters,
and the need for normalizing the bias of feature selection criteria with
respect to dimension.  We show theoretical proofs on the dimensionality
biases, and present a normalization scheme that can be applied to any
criteria to ameliorate these biases.  In addition to our automated
algorithm, we developed Visual-FSSEM, which incorporates visualization and
feature selection in an interactive environment.  Finally, we present an
automated approach for building hierarchical mixtures of probabilistic
principal component analyzers.  But, before I talk about these various
algorithms, I'm going to talk about an application area where I first
encountered the need to perform feature selection and clustering
simultaneously, which is medical content-based image retrieval.

Host: Simon Kasif (sullivan.bu.edu/kasif)