Department of Statistics & Operations Research
Statistics Seminars
2009/2010
To subscribe to the list, please follow this link or send email to 12345saharon@post.tau.ac.il54321 (remove numbers unless you are a spammer…)
Second semester
|
16 February |
Yaacov Ritov, Hebrew University |
|
Early bird
special |
|
|
16 March |
Ruth Heller, Technion |
|
|
Sensitivity analysis for the cross-match test, with applications
in genomics |
|
23 March |
Ishay
Weissman, Technion |
|
|
Dependence Measures for Multivariate Extreme Value
Distributions |
|
13 April |
Philip Stark, |
|
|
Justice and Inequalities |
|
27 April |
Bradly
Jones, JMP Software |
|
|
|
|
4 May |
Yoram
Gal-Ezer |
|
|
Where was Bonferroni logically wrong
and does the FDR correct his mistake? |
|
11 May |
Boaz Nadler, Weizmann
Institute |
|
|
Principal Component Analysis in Noisy High Dimensional Settings |
|
20 May |
Jason Fine, |
|
Special Thursday time |
|
|
27 May |
Yair
Goldberg, |
|
Special Thursday time |
Censored quantile regression using inverse
probability of censoring weighted average |
|
1 June |
|
|
|
First Semester
|
3 November |
Ronny Luss, Tel Aviv University |
|
|
Predicting Abnormal Returns From News Using Text Classification |
|
10 November |
|
|
|
|
|
24 November |
Elad Hazan, IBM Research |
|
|
|
|
15 December |
Nayantara
Bhatnagar, |
|
|
Rapid and Slow Convergence of Simulated Tempering and
Swapping |
|
22 December |
Yuval Nardi,
Technion |
|
11:00 am |
Maxima of asymptotically Gaussian random fields ***(note special
time)*** |
|
29 December |
Alan Izenman,
|
|
|
Regularization, Sparsity, and Rank
Restrictions in High-Dimensional Regression |
|
5 January |
Gal Elidan, Hebrew University |
|
|
|
|
19 January |
Malka Gorfine, Technion |
|
|
Statistical Methods for Genetic Risk Estimation of Rare Complex
Genetic Diseases |
Seminars are held on Tuesdays, 10.30 am, Schreiber Building, 309 (see the TAU map ). The seminar organizer is Saharon Rosset.
To join the seminar mailing list or any other inquiries - please call (03)-6408820 or email 12345saharon@post.tau.ac.il54321 (remove numbers unless you are a spammer…)
Seminars from previous years
- 2008-2009
- 2007-2008 (organized
by
- 2006-2007
(organized
by
- 2005-2006
(organized
by
- 2004-2005
(organized
by
ABSTRACTS
- Ronny Luss,
Predicting
Abnormal Returns From News Using Text Classification
Abstract:
We show how text from news articles can be used to predict intraday price movements
of financial assets using support vector machines. Multiple kernel learning is
used to combine equity returns with text as predictive features to increase
classification performance and we develop an analytic center cutting plane
method to solve the kernel learning problem efficiently. We observe that while
the direction of returns is not predictable using either text or returns, their
size is, with text features producing significantly better performance than
historical returns alone.
Can we infer
historical population movements from principal component analysis of genetic
data? A 30-year old argument rages on
Abstract:
The seminal Science paper by Menozzi, Piazza and Cavalli-Sforza in 1978, and the book by the same authors in
1994, established the use of principal component analysis of genetic data for
making inferences about human history and migration. Specifically, the 1978
paper concluded that the Neolithic expansion (circa 6000 BC) had a major effect
on the European genetic landscape. In 2008, a Nature Genetics paper by Novembre and Stephens claimed that the apparent patterns in
these original works "resemble mathematical artifacts" which are
expected if the genetic data were generated by local gene exchange only (i.e.,
no long-range migration). Their arguments are based on the properties of Toeplitz matrices and their eigen-decompositions.
We re-examine the properties of the original data and the relevant mathematical
results, and demonstrate that the arguments of Novembre
and Stephens do not apply in this case. We also perform a critical re-analysis
of the original data with "modern" tools and conclude that the
original conclusions are statistically valid, though their historical
interpretation is difficult to verify.
Decision-making
under uncertainty for structured problems
Abstract:
Decision-making in the face of uncertainty over future outcomes is a
fundamental problem of statistics and operations research with numerous
applications. In this talk I'll describe recent algorithmic advances, both in
terms of accuracy as well as computational efficiency.
We describe the first efficient algorithm for the problem of online linear
optimization in the limited-feedback (bandit) setting which achieves the
optimal regret bound. This resolves an open question since the work of Awerbuch and Kleinberg in 2004, and is made possible via a
new technique for controlling the exploration-exploitation tradeoff, inspired
by convex optimization. Next we describe new prediction algorithms which attain
optimal regret bounds in both worst case and stochastic scenarios. Tight
performance bounds for prediction which interpolate between the worst-case and
stochastic approaches were considered a fundamental open question.
Based on work with Jacob Abernethy, Satyen Kale and
Alexander Rakhlin.
Rapid and Slow
Convergence of Simulated Tempering and Swapping
Abstract: Markov
Chain Monte Carlo samplers are ubiquitous in statistical mechanics and Bayesian
statistics and have been analyzed extensively in theoretical computer science.
When the distribution being sampled from is multimodal, these samplers often
require a long running time to converge close to the desired distribution.
Multimodal posterior distributions arise very commonly in model selection,
mixture models and in statistical mechanical models. Simulated tempering and
swapping are two methods designed to sample more effectively from multimodal
distributions. In this work we show that
even these algorithms can fail to converge quickly and propose
modifications that can speed up the convergence.
- Yuval Nardi, Technion
Maxima of
asymptotically Gaussian random fields
Abstract:
The distribution of maxima of asymptotically (in a sense to be made precise in
the talk) Gaussian random fields over nice Euclidean sets is investigated. I
will describe a novel approach that may be used to yield asymptotic expansions
for such extremal probabilities. The approach builds
up on a measure transformation argument followed by some local approximation
arguments. A specific application from the realm of signal detection will
accompany the derivation. If time permits, I will show how to utilize the
approach for constructing simultaneous confidence bands for an unknown
(multivariate) density function.
- Alan J. Izenman,
Regularization, Sparsity, and Rank Restrictions in High-Dimensional
Regression
Abstract: As
enormous data sets become the norm rather than the exception, statistics as a
scientific discipline is changing to keep up with this development. Of particular interest are regression
problems in which attention to high dimensionality has become an important part
in determining how to proceed. In
multiple regression, regularization and sparsity
considerations have led to new methodologies for dealing with the
high-dimensionality, low sample-size situation.
In multivariate regression, rank restrictions have led to a reduced-rank
regression model that incorporates many of the classical
dimensionality-reduction methodologies, such as principal component analysis
and canonical variate analysis, as special
cases. In this talk, we discuss problems
of working with regression data when there are a large number of variables and
a relatively small number of observations, and we explore some new graphical
ideas for determining the effective dimensionality of multivariate regression
data.
- Gal Elidan,
The "Ideal
Parent" Algorithm
Abstract:
Bayesian networks are a formalism for encoding high-dimensional structured
joint distributions. The appeal of Bayesian networks is that an intuitive graphical representation
combined with a principled probabilistic
foundation lead to a compact representation of the distribution in a
decomposable form. This compact representation also facilitates efficient
methods for performing probabilistic computations, and automatic methods for parameter
estimation. Indeed, the past two decades
have seen an exponential growth in research related to these models.
Despite many innovative advances, model selection, or searching for a
beneficial structure of a Bayesian network, remains a formidable computational
task, which limits most applications to parameter estimation. This problem is
even more acute when learning networks in the presence of missing values or
hidden variables --- a scenario that is part of many real-life problems.
In this work we present a general method for dramatically
speeding model selection for continuous
variable Bayesian networks with common parametric distributions. In short, we efficiently evaluate the
approximate merit of candidate structure
modifications and apply time consuming (exact) computations only to
the most promising ones, thereby
achieving significant improvement in the running time of the search algorithm, without
compromising the quality of the solution.
Our method also naturally and efficiently facilitates the addition of
useful new hidden variables into the
network structure --- an automatic factor analysis like task that is typically considered both conceptually
difficult and computationally prohibitive. We demonstrate our method on
synthetic and real-life datasets, both for learning structure on fully and partially observable data, and for
introducing new hidden variables during structure search.
Statistical
Methods for Genetic Risk Estimation of Rare Complex Genetic Diseases
Abstract:
With the advances in the genetic dissection of complex diseases, the public has
been increasingly interested in an individual's genetic risk for developing
these diseases. Generally, there are two
aspects to the estimation of genetic risk: estimation of mutation carriership probability for a disease gene and prediction
of disease probability given the mutation status of the disease gene. Residual
risk heterogeneity widely exists even after adjusting for the disease gene and
thus it is important for obtaining accurate risk estimation. However, residual
risk heterogeneity is being ignored in all the current available estimation
procedures. We propose to account for the residual risk heterogeneity through
the use of frailty models and data from case-control family study. Another
common complication in complex diseases is that a disease gene can affect
multiple diseases. Thus, a subject censored due to another cause that is
related to the same gene is no longer independent of the age at onset of the
primary disease under study. We tackle this problem in the competing risks
framework. All the new estimation procedures developed in this work are
investigated extensively by simulations, and their asymptotic properties are
provided. The methods are illustrated with real data sets.
Lost with a MAP
in a foreign neighborhood
Abstract: We consider the maximal a-posteriori path (MAP)
estimator of an HMM process. We show that this estimator may be unreasonable
when the state space is non-finite, or the process is in continuous time. We
argue that this sheds a doubt on the usefulness of the concept in the standard
finite state space in discrete time HMM model. We then discuss a similar
phenomena in the completely different model of sparse regression.
- Ruth Heller, Technion
Sensitivity
analysis for the cross-match test, with applications in genomics
Abstract: The cross-match test is an exact, distribution
free test of no treatment effect on a high dimensional outcome in a randomized
experiment. The test uses optimal nonbipartite
matching to pair 2I subjects into I pairs based on similar outcomes, and the
cross-match statistic A is the number of times a treated subject was paired
with a control, rejecting for small values of A. If the test is applied in an
observational study in which treatments are not randomly assigned, it may be
comparing treated and control subjects who are not comparable, and may
therefore falsely reject a true null hypothesis of no treatment effect. We
develop a sensitivity analysis for the cross-match test, and apply it in an
observational study of the effects of smoking on gene expression levels. In
addition, we develop a sensitivity analysis for a standard multiple testing
procedure using the cross-match test and apply it to 1762 molecular function
categories in Gene Ontology.
Based on work with Shane Jensen, Paul Rosenbaum, and Dylan Small.
- Ishay
Weissman, Technion
Dependence
Measures for Multivariate Extreme Value Distributions
Abstract: The dependence structure of multivariate extremes
will be discussed first. Then, two dependence measures will be presented. These
measures are suitable for any number of dimensions and are invariant under
increasing transformations of the components. They possess an additional
desired property, lacked by their competitors, which makes them natural
dependence measures for multivariate extremes. A surprising connection to the
largest spacing among iid uniform random variables
will be discussed. This connection is useful as a diagnostic tool for the
quality of random number generators.
- Philip
Stark,
Justice and
Inequalities
Abstract: I will discuss some problems in election auditing
and litigation that can be solved using probability inequalities. The lead example, illustrated with case
studies in auditing elections and estimating damages in civil litigation, is to
construct nonparametric one-sided confidence bounds for the mean of a
nonnegative population. If time permits, I will also discuss a contested
election in which a simple probability inequality provided evidence the court
found persuasive. This seminar is partly
a plea for help from probabilists: I hope someone in the audience can point me
to inequalities that are sharper than those I'm using.
Efficient Designs with Minimal Aliasing
Abstract: For some experimenters, a disadvantage of the
standard optimal design approach is that it does not consider explicitly the
aliasing of specified model terms with terms that are potentially important but
are not included in the model. For example, when constructing an optimal design
for a first-order model, aliasing of main effects and interactions is not
considered. This can lead to designs that are optimal for estimation of the
primary effects of interest, yet have undesirable aliasing structures. In this
talk, I explain how to construct exact designs that minimize expected squared
bias subject to constraints on design efficiency. I will demonstrate use of the
method using several examples that allow for comparison with standard textbook
approaches.
Where was Bonferroni logically wrong and does the FDR correct his mistake?
Abstract: In cases of a first positive result after many
tries, with large enough sample sizes to detect a considerable effect, the need
to correct P value is quite intuitive. But this need is really justified not
for the increased chance to encounter a false positive result as usually
thought, whereas the chance to encounter a real positive is also increased. The
real reason is to make up for a low prior expectation for a real effect in the
specific positive result, because the absence of significance in the other
tries probably came from a low prior probability.
This pitfall apparently mislead Bonferroni
and made him offer a wrong formula for the
adjustment of P value. It was wrong because it did not address the
mentioned effect of prior expectation
for a real effect. The only approach completely capable of handling prior
expectation is Bayes's. The only role of a 'family'
of multiple comparisons is to serve as a data base for the assessment of prior
expectation. For this task the appropriate comparisons to be included are as
many as available comparisons of assumed equal prior probability for a real effect.
The FDR approach will be shown to be actually located half
the way to a completely prior expectation oriented approach. But this is not
really enough for assessing the credibilty of
specific finding. The thought it can be used for this purpose will be presented
as an "optical illusion".
In any case that either such "data base" for the
assessment of the prior expectation, or general guidelines for the prior are
available, it is not just a matter of choice to take it into account. Any other
option will bring to much less realistic results, unless a uniform distribution
of the prior can be considered no less realistic than other distributions.
Principal
Component Analysis in Noisy High Dimensional Settings
Abstract: Principal Component Analysis (PCA) is perhaps the
most widely used method in multivariate analysis.
In this talk I'll first review some recent results regarding the behavior of
the first few largest eigenvalues and eigenvectors of
PCA when the observed high dimensional data is of low rank but corrupted by
noise.
Second, I'll present some applications of these results, mainly to the problem
of non-parametric detection of signals embedded in noise.
- Jason
Fine,