Latent Tree Models for Density Estimation: Algorithms and Applications

PhD Thesis Proposal Defence


Title: "Latent Tree Models for Density Estimation: Algorithms and Applications"

by

Mr. Yi WANG


Abstract:

Latent tree models are tree-structured Bayesian networks in which
variables at leaf nodes are observed and called manifest variables,
while variables at internal nodes are unobserved and called latent
variables. Latent tree models can represent complex relationships among
manifest variables. In the meantime, inference in such models is
computationally simple.

In this proposal, we investigate the usefulness of latent tree models to
density estimation, and explore its applications to (1) classification
and (2) approximate inference in general Bayesian networks. In the first
application, we learn a latent tree model to estimate each
class-conditional distribution of attributes, and use Bayes rule to make
prediction. The outcome is the so called latent tree classifier that has
low classification error. It can also reveal underlying concepts and
discover interesting subgroups within each class. In the second
application, we build a latent tree model offline to approximate the
joint probability distribution represented by a Bayesian network, and
when online, make inference with the latent tree model instead of the
original Bayesian network. Research on this thread leads to a novel
approximate inference method that achieves good accuracy at low online
computational cost. To make those happen, we develop a learning
algorithm for each application. We also conduct empirical studies to
validate the values of the proposed approaches.


Date:     		Tuesday, 13 January 2009

Time:                   2:00p.m.-4:00p.m.

Venue:                  Room 3501
 			lifts 25-26

Committee Members:      Dr. Nevin Zhang (Supervisor)
 			Dr. Brian Mak (Chairperson)
                         Dr. Lei Chen
 			Prof. Dit-Yan Yeung


**** ALL are Welcome ****