|
Hui Zou talks with
ScienceWatch.com and answers a few questions about
this month's New Hot Paper in the field of
Mathematics.
|
|
Article Title: The adaptive lasso and its oracle
properties
Authors:
Zou, H
Journal: J AMER STATIST ASSN
Volume: 101
Issue: 476
Page: 1418-1429
Year: DEC 2006
* Univ Minnesota, Sch Stat, Minneapolis, MN 55455
USA.
* Univ Minnesota, Sch Stat, Minneapolis, MN 55455 USA.
|
|
Why do you think your paper is highly
cited?
This paper reveals some fundamental properties of the L1 (lasso) method
which has been, during the last six years, one of the most popular topics
in statistics, machine learning, and signal processing. It proves some
limitations of the classical lasso and provides an improved version named
the “adaptive lasso.” This paper also connects the L1
penalization and the non-concave penalization for the first time. The later
is another “hot” model selection method.
Does it describe a new discovery, methodology, or
synthesis of knowledge?
|
"The current
variable selection methods often
assume some strong model
structures. We need new methodology
and theory to help remove such
rigid assumptions."
|
|
We closely examined the theoretical properties of the classical lasso and
discovered some fundamental limitations. We then further proposed a new
method, named the adaptive lasso, which can remove all the limitations by a
simple modification on the classical lasso.
Would you summarize the significance of your paper
in layman’s terms?
It is critically important to have an efficient and powerful method for
feature selection in order to extract useful knowledge from massive
high-dimensional data. We invented a method that is theoretically sound and
computationally efficient for performing feature selection with
high-dimensional data.
How did you become involved in this research, and
were there any problems along the way?
Our research was partly motivated by a conjecture raised by Jianqing Fan
and Runze Li, in their article: “Variable selection via nonconcave
penalized likelihood and its. oracle properties” (J. Amer.
Statist. Assoc. [96]: 1348-1360, 2001.) One contribution of our paper
is that we rigorously proved their conjecture.
Where do you see your research leading in the
future?
The current variable selection methods often assume some strong model
structures. We need new methodology and theory to help remove such rigid
assumptions.
Do you foresee any social or political
implications for your research?
Conceptually, our method should be able to be applied to all data-analytic
problems where variable selection is needed. For instance, in many medical
studies, researchers want to discover new bio-markers or identify a set of
genes that may cause a specific disease.
Hui Zou, Assistant Professor
School of Statistics
University of Minnesota
Minneapolis, MN, USA
