题目：DiscLDA: Discriminative Learning for Dimensionality Reductionand Classif
ication 时间：12月25日 10:30-11:30
地点：蒙民伟楼404会议室
报告人：Fei Sha
Assistant Professor
Department of Computer Science, University of Southern California, USA
摘要：
Probabilistic topic models have become popular as methods for dimensionality r
eduction in collections of text documents or images. These models are usually
treated as generative models and trained using maximum likelihood or Bayesian
methods. In this talk, we discuss an alternative: a discriminative framework i
n which we assume that supervised side information is present, and in which we
wish to take that side information into account in finding a reduced dimensi
onality representation. Specifically, we present DiscLDA, a discriminative va
riation on Latent Dirichlet Allocation (LDA) in which a class-dependent linear
transformation is introduced on the topic mixture proportions. This parameter
is estimated by maximizing the conditional likelihood. By using the transfor
med topic mixture proportions as a new representation of documents, we obtain
a supervised dimensionality reduction algorithm that uncovers the latent struc
ture in a document collection while preserving predictive power for the task o
f classification. We compare the predictive power of the latent structure of D
iscLDA with unsupervised LDA on the 20 Newsgroups document classification task
and show how our model can identify shared topics across classes as well as c
lass-dependent topics.
Joint work with Simon Lacoste-Jullien (UC Berkeley) and Michael I.Jordan (UC B
erkeley)
简历：
Fei Sha got his Ph.D from U. of Pennsylvania in computer and information scien
ce. Afterwards, he spent a year at UC Berkeley as a postdoc with Prof. Michae
l Jordan and Prof. Stuart Russell. He then joined Yahoo Research as a research
scientist for a year. He has been a faculty at USC's computer science departm
ent since Aug., 2008. His research interest focuses on statistical machine lea
rning. His work had won best student paper awards at NIPS and ICML.
|
