UTCS Colloquia - Seyoung Kim/Carnegie Mellon University, "Lasso-type methods for multi-task regression with structured sparsity", ACES 2.402

Contact Name: 
Jenna Whitney
Mar 8, 2011 11:00am - 12:00pm

There is a sign-up schedule for this event that can be found at



Type of Talk: UTCS Colloquia

Speaker/Affiliation: Seyoung K

im/Carnegie Mellon University

Talk Audience: Faculty and Graduate Stud


Date/Time: Tuesday, March 8, 2011, 11:00 a.m.

Location: A

CES 2.402

Host: Pradeep Ravikumar

Talk Title: Lasso-type methods

for multi-task regression with structured sparsity

Talk Abstract:
this talk, I will consider the problem of learning a multi-task regressio

n with structured sparsity, where the multiple outputs have a complex corr

elation structure and the highly correlated outputs tend to share common re

levant input variables. Assuming that the correlation structure in the outp

uts is given as a graph or tree, we introduce two different lasso-type met

hods with novel penalty functions that are constructed based on the availab

le output correlation structure to encourage a shared sparsity pattern amon

g correlated outputs. For the case of graph-structured outputs, we use a f

usion penalty as a building block in our graph-guided fused lasso, whereas
for tree-structured outputs, we use a group-lasso-like penalty with weigh

ted overlapping groups in tree-guided group lasso. The optimization in thes

e methods is challenging because of the non-separability of the parameters

in the penalty functions. We describe a simple yet efficient proximal gradi

ent method that can be also used to solve general fused lasso or group lass

o with overlapping groups. I will discuss experimental results from applyin

g our methods to genetic association analysis.

Speaker Bio:

Kim is an assistant professor at the Lane Center for Computational Biology

, School of Computer Science, Carnegie Mellon University. She received her
B.S. in computer engineering from Seoul National University, Korea, Ph.

D. in Computer Science from the University of California, Irvine, and was
a postdoctoral fellow in the Machine Learning Department at Carnegie Mello

n University.