Belief Network Induction
Ron Musick
Computer Science Division
University of California
Berkeley, CA 94720
Musick@cs.berkeley.edu
Abstract
This dissertation describes BNI (Belief Network Inductor), a tool that
automatically induces a belief network from a database. The
fundamental thrust of this research program has been to provide a
theoretically sound method of inducing a model from data, and
performing inference over that model. Along with a solid grounding in
probability theory, BNI has proven to be a quick, practical method of
inducing data models that are highly accurate. The results include a
belief network that stores beta distributions in the conditional
probability tables, coupled with theorems demonstrating how to
maintain these distributions through inference; techniques for
applying neural network and other learning techniques to the task of
conditional probability table learning; and a decision theoretic
sampling theory which addresses scalability issues by characterizing
the size of the sample needed to produce high quality inferences.
The setting for this work is in database mining. Database mining is
one of the fastest growing topics in Artificial Intelligence today,
with industry providing at least as much impetus as research labs and
universities. The general goal is to extract interesting quantities
or relationships that are ``hidden'' in large corporate or scientific
databases, with the potential benefits of a successful technology
being enormous. For example, models can be built that characterize
what types of customers will respond to what types of marketing schemes,
retailers will be able to predict sales to help determine correct
inventory levels and distribution schedules, and insurance companies
will be able to predict expected claim costs and better classify who
will buy what type of coverage.
Look at the Thesis (ps.gz, pdf.gz)