AUTHOR:

Thomas Hofmann 
Artificial Intelligence Laboratory
Center for Biological and Computational Learning
M.I.T.


TITLE:

Structuring Document Databases by Hierarchical Clustering and
Abstraction


ABSTRACT:

Statistical mixture models are versatile tools for many tasks in
machine learning, exploratory data analysis, and pattern
recognition. They are suitable to detect structures such as clusters
and data hierarchies and have a sound foundation in probability
theory. In this talk, I will present a novel hierarchical mixture
model for co-occurrence data and demonstrate its benefits in the
domain of information retrieval. The described learning architecture
generates hierarchical organizations of document databases from word
occurrence statistics. The key feature of the model is the combination
of the document hierarchy with an abstractive organization of
keywords. This supports the identification of discriminant terms and
has proven to be very useful for interactive retrieval, as well as for
problems like cluster summarization. To train the model I propose an
annealed variant of the Expectation-Maximization (EM) algorithm for
maximum likelihood estimation.