James G. Scott
University of Texas at Austin
1 University Station, B6500
Austin, TX 78712
Office: CBA 6.478
Phone: (512) 471-5905
Fax: (512) 471-0587
E-mail: james.scott@mccombs.utexas.edu
I am an assistant professor of statistics at UT-Austin. My appointment is held jointly in:
-the statistics group inside the Department of Information, Risk, and Operations Management in the McCombs School of Business.
-the Division of Statistics and Scientific Computation, an interdisciplinary unit housed in the College of Natural Sciences.
This fall (2011), I am coordinating the UT statistics seminar series, which is sponsored jointly by UT SSC, the Risk Analysis group of the McCombs School, and UT’s Intelligent Data Exploration and Analysis Laboratory (in conjunction with Yahoo!).
About my research
I am a Bayesian statistician (papers here). Much of my methodological research centers around the core issues of model choice, multiple testing, variable selection, and latent-feature extraction. I pay particular attention to issues of computation, scalability, and the quantification of uncertainty in these problems. My collaborative projects involve the application of Bayesian methodology to extract new insights from large data sets in neuroscience, clinical bio-informatics, astronomy, finance, politics, corporate strategy, and higher education.
Most recently I have been studying connections between statistical machine learning and Bayesian shrinkage estimation; variable selection and forecasting in non-linear, non-Gaussian models; regularization in high-dimensional contingency tables and mixed-membership models; and highly structured, dynamic models for correlation among continuous, ordinal, and multinomial outcomes.
I also study “scalable Bayes”—that is, tools that allow Bayesian-inspired approaches for sparse-signal detection to be applied on enormous data sets, where many traditional tools simply won't work. I avoid all-purpose algorithms, and instead try to exploit the probabilistic and mathematical structure of a problem in order to build tailored, parallelizable algorithms for specific situations (e.g. multinomial regression with thousands of categories and hundreds of millions of observations). Some of my current work on this topic involves the theory of random matrix projections to create new algorithms for sparse-signal detection in non-Gaussian, non-convex models. I believe that these methods offer a promising alternative to traditional coordinate-descent or gradient-based methods.
About my courses
I teach STA 371, which is an upper-division undergraduate course on regression, probability models, and decision analysis for undergraduates at the McCombs School of Business. This snippet, taken from the introduction to the class notes, may give you an idea of what the course is like.
In Spring 2012, I will also teach SSC 325H, an introduction to data-based reasoning and statistical inference for students in any undergraduate honors program at UT. The only pre-requisite is a course on differential and integral calculus.
About me
I grew up in Katy, Texas (before this–eek!–was there); majored in math and liberal arts (Plan II) at UT-Austin; studied for a master’s in mathematics at Cambridge University (Trinity College); and earned my Ph.D. in statistics from Duke University in May of 2009 under the supervision of Jim Berger. I have been on the faculty here at UT since July of 2009.