Speaker/Affiliation: Tamar

a G. Kolda/Sandia Labs

Date/Time: Thursday August 28 2008 3:30 p

.m.

Location: ACES 6.304

Host: Inderjit Dhillon

Talk
Title: Tensor Decompositions and Data Mining

Talk Abstract:
Mat

rix decompositions such as the singular value
decomposition (SVD) are w

ell-known tools for data
mining that have been used in a wide variety o

f
applications such as text retrieval web page ranking
and face r

ecognition. But matrix decompositions are
restricted to two-way tabula

r data. In many cases it
is more natural to arrange data into an N-way
array
where N > 2 which is referred to as a higher-order
tensor.

I will describe tensor decompositions such as
CANDECOMP/PARAFAC and Tuc

ker. These
decompositions have been in use for several decades
in p

sychometrics and chemometrics and have recently
become popular in signa

l processing numerical analysis
and many other fields. I will talk ab

out computational issues
and Sandia''s Tensor Toolbox for MATLAB. I wil

l present
several examples of tensor decompositions being used
for

hyperlink analysis for web search computer vision
and neuroscience. I
will also mention other types of tensor
decompositions and some new ap