Title: Graph based Semi-supervised Learning and Spectral Kernel Design
Speaker: Tong Zhang
In this talk, I will discuss some general issues for constructing
discriminative learning models using semi-supervised learning (learning
methods using both labeled and unlabeled data).
I will then focus on recently popularized graph (aka, manifold) based
semi-supervised learning algorithms. I shall present a statistical
analysis which answers a number of previous unresolved questions.
A critical component of our analysis is an equivalence of graph-based
semi-supervised learning and supervised kernel learning, which we
establish first.
As an easy but important consequence, we prove the convergence of
graph methods when the number of unlabeled data increases.
We then show that graph based semi-supervised learning can be
viewed as a special case of a class of unsupervised kernel design
methods.
Using a transductive generalization bound on graphs, we can show that
good
unsupervised kernel design will be helpful and we will present the
optimal
kernel (based on the bound).
We will then use this result, and a statistical model to show why
graph-based methods are often effective.
Moreover, under the framework of unsupervised kernel design, I can
discuss
the limitations of graph methods, alternatives, and the
relationship of
several different methods.
Consequences of this analysis will be illustrated with empirical
experiments. This is a joint work with Rie Ando of IBM T.J. Watson
Research Center.
Please contant Prof. Peter Bartlett (bartlett at stat dot Berkeley dot
EDU) if you want
to join lunch with the speaker at noon.