Department of Mathematics and Statistics, Binghamton University seminars:sml

September 9, 2014

2017-06-12T11:43:59-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, September 9, 2014
Time: 12:00-1:00
Room: OW-100E
Speaker: Ganggang Xu (Mathematical Sciences)
Title: Asymptotic optimality and efficient computation of the leave-subject-out cross-validation

Abstract

Although the leave-subject-out cross-validation (CV) has been widely used in practice for tuning parameter selection for various nonparametric and semiparametric models of longitudinal data, its theoretical property is unknown and solving the associated optimization problem is computationally expensive, especially when there are multiple tuning parameters. In this paper, by focusing on the penalized spline method, we show that the leave-subject-out CV is optimal in the sense that it is asymptotically equivalent to the empirical squared error loss function minimization. An efficient Newton-type algorithm is developed to compute the penalty parameters that optimize the CV criterion. Simulated and real data are used to demonstrate the effectiveness of the leave-subject-out CV in selecting both the penalty parameters and the working correlation matrix.

September 23, 2014

2017-06-12T11:45:00-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, September 23, 2014
Time: 12:00-1:00
Room: OW-100E
Speaker: Wei Sun (Purdue University)
Title: Causal Inference Framework for Complex Advertising Effect Measurement

Abstract

As the online advertising industry has evolved into an age of diverse ad formats and delivery channels, users are exposed to complex ad treatments involving various ad characteristics. The diversity and generality of ad treatments call for accurate and causal measurement of ad effectiveness. In this talk, I will present a new causal inference framework for assessing the impact of general advertising treatments. It enables analysis on multi-dimensional ad treatments, where each ad treatment could be discrete or continuous. Our approach is unbiased and computationally efficient by employing a tree structure that specifies the relationship between user characteristics and the corresponding ad treatment. This framework is robust to model misspecification and highly flexible with minimal manual tuning. The efficacy of our approach is demonstrated in several advertising campaigns:

Investigate best frequency of ads in the campaign. We show that the ad frequency usually has a treatment effect cap and is often over-estimated by naive estimation. This investigation suggests adjusting the cap of the ad frequency to avoid wasting of ad inventory.
Investigate best ads size on different mobile devices. Our model suggests that the ad size 300*250 produces consistently high success rates across various mobile devices.

Bio: Wei Sun is a graduate student from Department of Statistics at Purdue University. His research interests expand various aspects of statistical machine learning and high dimensional data analysis. He has published in Journal of Machine Learning Research and Electronic Journal of Statistics. Wei Sun received an Institute of Mathematical Statistics Travel Award in 2013.

Itinerary
9:30 - 10:00	Xingye Qiao
10:00 - 10:30	Graduate student(s)
10:30 - 11:00	Qiqing Yu
11:00 - 11:30	Ganggang Xu
12:00 - 1:00	Talk
1:00 - 2:00	Lunch with graduate students
2:00	Departure

November 18, 2014

2014-11-06T22:39:43-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, November 18, 2014
Time: 12:00-1:00
Room: OW-100E
Speaker: Ming Yang (Computer Science)
Title: Multi-task Learning with Gaussian Matrix Generalized Inverse Gaussian Model

Abstract

In this talk, we first introduce the multi-task learning problem and the application of Matrix Generalized Inverse Gaussian (MGIG) distribution to the problem. We still propose the GMGIG regression model for multi-task learning. To make the computation tractable, we simultaneously use variational inference and sampling techniques. In particular, we propose two sampling strategies for computing the statistics of the MGIG distribution.

December 2, 2014

2014-11-05T23:20:29-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, December 2, 2014
Time: 12:00-1:00
Room: OW-100E
Speaker: Yingming Li (Computer Science)
Title: Scientific Articles Recommendation

Abstract

Advances in social network sites enable researchers to access to large on-line archives of scientific articles. These social network sites allow researchers to create their personal libraries for the on-line articles that interest them and to share the libraries with other researchers. This makes recommender systems helpful for researchers to find interesting articles. In this paper, we propose topic regression matrix factorization models to recommend scientific articles for on-line community. The main idea of topic regression Matrix Factorization models lies in extending the matrix factorization with a probabilistic topic modeling. Further, we demonstrate the efficacy of topic regression Matrix Factorization models on a large subset of the data from CiteULike, a bibliography sharing service dataset.

September 29, 2015

2015-09-11T14:55:45-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, September 29, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Xu Chu (Mathematical Sciences)
Title: Lasso, AdaLasso and Sure Independence Screening

Abstract

This is the first of a series of three talks which focus on variable selection for high-dimensional data. I will introduce the following three papers on variable selection.

Tibshirani, R. (1996) Regression shrinkage and selection via the lasso. J. R. Statist. Soc. B, 58, 267–288.
Zou, H. (2006) The adaptive lasso and its oracle properties. J. Am. Statist. Ass., 101, 1418–1429.
Fan, J. and Lv, J. (2008) Sure independence screening for ultrahigh dimensional feature space (with discussion). J. R. Statist. Soc. B,70, 849–911.

October 06, 2015

2015-09-11T14:57:22-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, October 06, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Wenbo Wang (Mathematical Sciences)
Title: Selection consistency of lasso and Stability Selection

Abstract

This is the second of a series of three talks which focus on variable selection for high-dimensional data. I will introduce the following three papers on variable selection, especially on the selection consistency.

Meinshausen, N. and Bühlmann, P. (2006) High dimensional graphs and variable selection with the lasso. Ann. Statist., 34, 1436–1462.
Zhao, P. and Yu, B. (2006) On model selection consistency of lasso. J. Mach. Learn. Res., 7, 2541–2563.
Meinshausen, N., & Bühlmann, P. (2010). Stability selection. J. R. Statist. Soc. B, 72(4), 417-473.

October 13, 2015

2015-09-11T15:00:55-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, October 13, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Wenming Deng (Mathematical Sciences)
Title: The Bolasso, the Dantzig selector and more on Stability Selection

Abstract

This is the third and final talk of a series of three talks which focus on variable selection for high-dimensional data. I will introduce the two related work below, and then discuss more on Meinshausen & Bühlmann (2010) following the talk last week.

Bach, F. (2008) Bolasso: model consistent Lasso estimation through the bootstrap. In Proc. 25th Int. Conf. Machine Learning, pp.33–40. New York: Association for Computing Machinery.
Candes, E. and Tao, T. (2007) The Dantzig selector: statistical estimation when p is much larger than n. Ann. Statist., 35, 2312–2351.

October 20, 2015

2015-10-07T22:21:49-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, October 20, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Xin Qi (Mathematical Sciences)
Title: On False Discovery Rate (I)

Abstract

This is the first of three talks on false discovery rate on the topic of multiple testing. We will discuss the seminal papers, Benjamini and Hochberg (1995) and Storey (2002).

Benjamini, Y. and Hochberg, Y. (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society, Series B, 57, 289-300.
Storey, J.D. (2002). A Direct Approach to False Discovery Rates. Journal of the Royal Statistical Society, Series B, 64, 479-498.

October 27, 2015

2015-10-07T22:21:32-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, October 27, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Lin Yao (Mathematical Sciences)
Title: On False Discovery Rate (II)

Abstract

This is the second of three talks on false discovery rate on the topic of multiple testing. We will discuss the following two papers.

Storey, J.D., Taylor, J.E. and Siegmund, D. (2004). Strong Control, Conservative Point Estimation and Simultaneous Conservative Consistency of False Discovery Rates: A Unified Approach. Journal of the Royal Statistical Society, Series B, 66, 187-205.
Fan, J., Han, X., & Gu, W. (2012). Estimating false discovery proportion under arbitrary covariance dependence. Journal of the American Statistical Association, 107(499), 1019-1035.

November 3, 2015

2015-11-01T22:57:44-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, November 3, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Xiaojie Du (Mathematical Sciences)
Title: Inference About the Slope in Linear Regression with Missing Responses: An Empirical Likelihood Approach.

Abstract

This article considers linear regression models with responses that are allowed to be missing at random, which covers the model with fully observed data as a special case. We assume that the covariates and the errors are independent, without specifying their distributions. The main result of this article is to present two efficient maximum empirical likelihood estimators for the regression parameter, which are easy to obtain numerically. This fills a gap in the literature which does not provide a parameter estimator that is both simple and efficient: the usual efficient approaches require an estimator of the influence function, which can be quite involved. We also present an asymptotic confidence interval for the slope.

November 10, 2015

2015-09-11T15:02:56-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, November 10, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Junyi Dong (Mathematical Sciences)
Title: The Family of Distributions With Proportional Hazards

Abstract

The proportional hazards (PH) model specifies distributions for regression data. Given a covariate value, the PH model specifies a family of distributions for a random variable, called the PH family. I will talk about the method of generating pseudo random numbers under various special cases when the random variable may take infinity with probability one. I will also talk about the identifiability condition of the parameters in the PH family and the semi-parametric estimation of the parameters under the identifiability condition.

December 1, 2015

2015-11-02T17:07:04-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, December 1, 2015
Time: 12:00-1:00
Room: WH-100E
Speaker: Chen Liang (Mathematical Sciences)
Title: On Information Criterion

Abstract

I will present two papers both on the related concepts of information criterion.

Chen, J., & Chen, Z. (2008). Extended Bayesian information criteria for model selection with large model spaces. Biometrika, 95(3), 759-771.
Zhang, Y., Li, R., & Tsai, C. L. (2010). Regularization parameter selections via generalized information criterion. Journal of the American Statistical Association, 105(489), 312-323.

December 8, 2015

2015-11-23T16:55:51-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, December 8, 2015
Time: 12:00-1:00
Room: WH-G02
Speaker: Baiyang Qi (Mathematical Sciences)
Title: Building a model for scoring 20 or more runs in a baseball game

Abstract

I will present a paper on sports statistics. The abstract of that paper is as follows.

How often can we expect a Major League Baseball team to score at least 20 runs in a single game? Considered a rare event in baseball, the outcome of scoring at least 20 runs in a game has occurred 224 times during regular season games since 1901 in the American and National Leagues. Each outcome is modeled as a Poisson process; the time of occurrence of one of these events does not affect the next future occurrence. Using various distributions, probabilities of events are generated, goodness-of-fit tests are conducted, and predictions of future events are offered. The statistical package R is employed for analysis.

Febuary 23, 2016

2016-02-22T10:26:31-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, Febuary 23, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Xin Qi (Mathematical Sciences)
Title: RE: Joint estimation of multiple graphical models

Abstract

Gaussian graphical models explore dependence relationships between random variables, through the estimation of the corresponding inverse covariance matrices. In this paper we develop an estimator for such models appropriate for data from several graphical models that share the same variables and some of the dependence structure. In this setting, estimating a single graphical model would mask the underlying heterogeneity, while estimating separate models for each category does not take advantage of the common structure. We propose a method that jointly estimates the graphical models corresponding to the different categories present in the data, aiming to preserve the common structure, while allowing for differences between the categories. This is achieved through a hierarchical penalty that targets the removal of common zeros in the inverse covariance matrices across categories. We establish the asymptotic consistency and sparsity of the proposed estimator in the high-dimensional case, and illustrate its performance on a number of simulated networks. An application to learning semantic connections between terms from webpages collected from computer science departments is included.

March 1, 2016

2016-03-07T09:54:26-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, March 1, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Wenbo Wang (Mathematical Sciences)
Title: RE: Multicategory Angle-based Large-margin Classification and Reinforced SVM

Abstract

Large-margin classifiers are popular methods for classification. Among existing simultaneous multicategory large-margin classifiers, a common approach is to learn k different functions for a k-class problem with a sum-to-zero constraint. Such a formulation can be inefficient. We propose a new multicategory angle-based large-margin classification framework. The proposed angle-based classifiers consider a simplex-based prediction rule without the sum-to-zero constraint, and enjoy more efficient computation. Many binary large-margin classifiers can be naturally generalized for multicategory problems through the angle-based framework. Theoretical and numerical studies demonstrate the usefulness of the angle-based methods.

The Support Vector Machine (SVM) is a very popular classification tool with many successful applications. It was originally designed for binary problems with desirable theoretical properties. Although there exist various Multicategory SVM (MSVM) extensions in the literature, some challenges remain. In particular, most existing MSVMs make use of k classification functions for a k-class problem, and the corresponding optimization problems are typically handled by existing quadratic programming solvers. In this paper, we propose a new group of MSVMs, namely the Reinforced Angle-based MSVMs (RAMSVMs), using an angle-based prediction rule with k − 1 functions directly. We prove that RAMSVMs can enjoy Fisher consistency. Moreover, we show that the RAMSVM can be implemented using the very efficient coordinate descent algorithm on its dual problem. Numerical experiments demonstrate that our method is highly competitive in terms of computational speed, as well as classification prediction performance. Supplemental materials for the article are available online.

March 8, 2016

2016-03-07T09:56:31-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, March 8, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Lin Yao (Mathematical Sciences)
Title: RE: Sparse Regression Incorporating Graphical Structure among Predictors

Abstract

With the abundance of high dimensional data in various disciplines, sparse regularized techniques are very popular these days. In this paper, we make use of the structure information among predictors to improve sparse regression models. Typically, such structure information can be modeled by the connectivity of an undirected graph using all predictors as nodes of the graph. Most existing methods use this undirected graph edge-by-edge to encourage the regression coefficients of corresponding connected predictors to be similar. However, such methods do not directly utilize the neighborhood information of the graph. Furthermore, if there are more edges in the predictor graph, the corresponding regularization term will be more complicate. In this paper, we incorporate the graph information node-by-node, instead of edge-by-edge as used in most existing methods. Our proposed method is very general and it includes adaptive Lasso, group Lasso, and ridge regression as special cases. Both theoretical and numerical studies demonstrate the effectiveness of the proposed method for simultaneous estimation, prediction and model selection.

http://www.tandfonline.com/doi/abs/10.1080/01621459.2015.1034319

March 15, 2016

2016-03-07T10:07:06-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, March 15, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Junyi Dong (Mathematical Sciences)
Title: Diagnostic Plotting Methods for Proportional Hazards Models With Time-dependent Covariates or Time-varying Regression Coefficients

Abstract

This is the first one of a series of two talks, the second of which will be on Thursday.

Given a sample of regression data from $(Y,Z)$, we propose several new diagnostic plotting methods for the Cox models with the general external time-dependent covariates $Z$, which can be continuous, or with the time-varying regression coefficients. The main approach compares the non-parametric MLE of the survival function of $Y$ against its expected (correct) expression under the given Cox model which can be mis-specified. This is different from the existing methods in the literatures such as log-log plots and residuals plots. The simulation studies and the data analysis of our breast cancer data suggest that the combination of these diagnostic plotting methods we proposed performs quite satisfactorily. These new diagnostic plotting methods naturally yield various tests for checking the validity of the Cox models. The main advantage of the new tests over the residual tests is in the case that the data do not fit any PH model. Then the new tests are still valid but not the residual tests.

March 22, 2016

2016-02-28T21:26:30-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, March 22, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Grace Wang (Syracuse University)
Title: ConceFT: Concentration of Frequency and Time via a multitapered synchrosqueezed transform

Abstract

Time-frequency representations provide a powerful tool for the analysis of time series signals. Techniques that decompose the time-dependent signals into multiple oscillatory components, with time-varying amplitudes and instantaneous frequencies are very appealing and have been shown to be useful in a wide range of applications including geophysics, biology, medicine, finance and social dynamics. In this talk, I’ll give an introduction to time-frequency representations and review existing methods for the previously described decomposition. Then I’ll present a new method that applies the multitapering with synchrosqueezed transform. Numerical experiments as well as a theoretical analysis will be demonstrated to assess its effectiveness.

April 19, 2016

2016-04-08T12:56:34-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, April 12, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Ruiqi Liu (Mathematical Sciences)
Title: TBA

Abstract

Consider that we are observing iid copies $(X_i, Y_i)_{i=1}^n$ from random vector $(X, Y)$. According to some historical information, the marginal distributions of $X$ and $Y$ are known, but the joint distribution is unclear. A problem of interest is to estimate $\exp[h(X,Y)]$ for some measurable function $h$. This is of application value. For example, in insurance industry, some life insurance policies will cover both husband and wife . Let $X,Y$ be the left life time of husband and wife after signing the policy and $X, Y$ are usually dependent. The company is able to obtain the marginal distributions of $X$ and $Y$ from historical records. Often, the values of interest are $\min(X, Y)$, $\max(X, Y)$ or their distributions. This paper provides an empirical likelihood estimator to solve this problem. Some nice properties of our estimator are supported by theoretical analysis and simulation results.

April 19, 2016

2016-02-22T12:58:14-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, April 19, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Rachael Kline (Mathematical Sciences)
Title: RE: Joint Estimation of Multiple Precision Matrices with Common Structures

Abstract

Estimation of inverse covariance matrices, known as precision matrices, is important in various areas of statistical analysis. In this article, we consider estimation of multiple precision matrices sharing some common structures. In this setting, estimating each precision matrix separately can be suboptimal as it ignores potential common structures. This article proposes a new approach to parameterize each precision matrix as a sum of common and unique components and estimate multiple precision matrices in a constrained L1 minimization framework. We establish both estimation and selection consistency of the proposed estimator in the high dimensional setting. The proposed estimator achieves a faster convergence rate for the common structure in certain cases. Our numerical examples demonstrate that our new estimator can perform better than several existing methods in terms of the entropy loss and Frobenius loss. An application to a glioblastoma cancer data set reveals some interesting gene networks across multiple cancer subtypes.

April 26, 2016

2016-04-20T22:11:00-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, April 26, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Wolfgang Wefelmeyer (Universität zu Köln)
Title: Density estimators in regression models with errors in covariates

Abstract

In regression models $Y=r(X)+\varepsilon$ with $X$ and $\varepsilon$ independent, the density of the response $Y$ can be estimated by a convolution of (kernel) estimators for the densities of $r(X)$ and $\varepsilon$. The rate of this convolution estimator depends on the smoothness of the densities of $X$ and $\varepsilon$ and on the smoothness and local flatness of the regression function $r$. When we observe the covariates $X$ with measurement errors, $Z=X+\eta$, we need deconvolution estimators for the densities of $X$ and $\varepsilon$ and for $r$. This is joint work with Anton Schick and Ursula U. Müller.

May 10, 2016

2016-05-09T15:06:51-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, May 10, 2016
Time: 12:00-2:00p
Room: WH-100E
Speaker: Qinggang Diao (Mathematical Sciences)
Title: Cox proportional hazards model with time-dependent covariates

Abstract

This PhD dissertation is divided into four chapters, where right-censored (RC) data and interval- censored (IC) data under several different types of time dependent covariates assumptions will be discussed.

In Chapter 0, we will introduce some basic concepts and notations about survival analysis.

Chapter 1 reproduces the paper of Yu et al.(2015). In this chapter piecewise Cox models with right-censored data will be discussed. Piecewise Cox models are regression models that follow dif- ferent Cox models when restricted to different time intervals. We study a general class of piecewise Cox models that involve a single cut point so that there are two separate Cox models correspond- ing to the two time intervals created. We discuss the computation of the semi-parametric maximum likelihood estimates (SMLE) of the parameters, with right-censored data, and a simplified algorithm for the maximum partial likelihood estimates (MPLE). Simulation studies suggest that MPLE com- pares favorably with its SMLE counterpart, even though the SMLE is more efficient. To assess the appropriateness of the model assumption, we propose a simple diagnostic plotting method. This method will enable us to determine an appropriate cut point. We show that the results for the case of a single cut point can be extended to involving more than one cut point. Finally, we apply the methodology we have developed for piecewise Cox models to the survival analysis of a long-term breast cancer follow-up study on the prognostic significance of bone marrow micrometastasis. Our diagnostic plots suggest that it is appropriate to apply the piecewise Cox model to our data.

In Chapter 2, we consider the time-dependent covariates proportional hazards (TDCPH) model with interval-censored (IC) relapse times under the distribution-free set-up. The partial likelihood approach is not applicable for IC data, thus we use the full likelihood approach. It turns out that under the TDCPH model with IC data, the semi-parametric MLE (SMLE) of the covariate effect under the standard generalized likelihood is not unique and is not consistent. In fact, the parameter under the TDCPH model with IC data is not identifiable unless some stronger assumptions are imposed. We propose a modification to the likelihood function so that its SMLE is unique. We show that the parameter is identifiable under certain regularity conditions. Under the regularity assumptions, our simulation studies suggest that such an SMLE is consistent and we also give a rigorous proof of the consistency. We apply the method to our cancer relapse time data and conclude that the bone marrow micrometastasis does not have a significant prognostic factor.

In Chapter 3, we consider the semi-parametric estimation problem under the proportional haz- ards (PH) model with continuous time-dependent covariates and interval-censored data. We show that unlike the PH model with time-independent covariates, if the observable random vector takes on finitely many values, then the parameters in the model are not identifiable and there exist no consistent estimators of the parameters. We establish the identifiability condition for this issue. It provides a guideline for carrying out simulation studies and for the proof of consistency of certain semi-parametric estimators. Moreover, the naive extension of the generalized likelihood function does not lead to a consistent estimator. We propose two proper modifications of the generalized likelihood function, and they both yield consistent estimators. We also carry out simulation studies for these estimators. The covariate z(t) = u1(t > c)(t − c) will be discussed.

September 27, 2016

2016-09-13T16:52:55-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, September 27, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Wei Qian (Rochester Institute of Technology)
Title: Sufficient Dimension Reduction in High Dimension

Abstract

Sufficient dimension reduction (SDR) is known to be a powerful tool to achieve data reduction and data visualization in regression and classification problems. In this work, we study the high-dimensional SDR problems and propose a unified solution with regularization. Under the setting p»n, consistency results are investigated for important SDR methods. Special sparse structures of large predictor and error covariance are considered for potentially improved performance. In addition, the proposed approach is equipped with a new algorithm to efficiently solve the regularized objective functions and a data-driven procedure to determine structure dimension, without the need to inverse a large covariance matrix. Simulations and real data analysis are performed to demonstrate promising applications of our proposal in high-dimensional settings.

October 18, 2016

2016-10-06T23:36:51-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, October 18, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Yang Feng (Columbia University)
Title: Community detection with nodal information

Abstract

Discovering community structure is one of the fundamental issues in the study of networked data. Most existing community detection approaches take merely edge information as inputs, and deliver suboptimal results for networks with nodal covariates available. Regarding those networks, it is desirable to leverage covariates information for the improvement of detection accuracy. Towards this goal, we propose a flexible network model incorporating nodal signals, and develop likelihood-based inference methods. We will present a systematic study from both theoretical and practical aspects. Our theoretical analysis demonstrates favorable asymptotic properties of the proposed approach. We then derive practical algorithms for the search of the theoretical estimators. Numerical experiments show the effectiveness of our method in utilizing nodal information across a variety of simulated and real networked datasets.

October 25, 2016

2016-09-13T17:15:56-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, October 25, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Guan Yu (University at Buffalo)
Title: Supervised Learning Incorporating Graphical Structure among Predictors

Abstract

With the abundance of high dimensional data in various disciplines, regularization techniques are very popular these days. Despite the success of these techniques, some challenges remain. One challenge is the development of efficient methods incorporating structure information among predictors. Typically, the structure information among predictors can be modeled by the connectivity of an undirected graph using all predictors as nodes of the graph. In this talk, I will introduce an efficient regularization technique incorporating graphical structure information among predictors. Specifically, according to the undirected graph, we use a latent group lasso penalty to utilize the graph node-by-node. The predictors connected in the graph are encouraged to be selected jointly. This new regularization technique can be used for many supervised learning problems. For sparse regression, our new method using the proposed regularization technique includes adaptive Lasso, group Lasso, and ridge regression as special cases. Theoretical studies show that it enjoys model selection consistency and acquires tight finite sample bounds for estimation and prediction. For the multi-task learning problem, our proposed graph-guided multi-task method includes the popular l2,1-norm regularized multi-task learning method as a special case. Numerical studies using simulated datasets and the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset also demonstrate the effectiveness of the proposed methods.

November 1st, 2016

2016-10-22T13:22:58-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, November 1, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Yu Chen (ECE at Binghamton University)
Title: Enabling Smart Urban Surveillance at The Edge

Abstract

The unprecedented urbanization and the staggering development of modern information technologies (IT) make Smart City attractive and achievable. Beyond the scope of traditional city services and applications, Smart Cities provide urban planners and policy makers proactively and timely information to obtain a dynamic and comprehensive understanding about the rhythm of our cities. Although Cloud Computing is considered the ideal platform for vast volume urban data storage and processing, the sustainability of Smart Cities necessitates the capability of computations and data analysis at the edge of the networks, especially for mission critical applications requiring real-time information fusion and on-site decision making. Fog Computing, an extension of Cloud Computing, enables heterogeneous mobile and smart computing devices at the edge to collaborate for instant decision making. In this talk, a smart urban surveillance platform will be introduced. Leveraging the underlay fog network, the real-time multi-target tracking task is accomplished. Compared with the Cloud Computing, the experimental results are very encouraging and validate the feasibility of smart urban surveillance for instant decision making using Fog Computing at the network edges.

November 8th, 2016

2016-11-06T21:17:45-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, November 8, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Chengbin Deng (Geography at Binghamton University)
Title: Use of Big Geospatial Data for Better Mapping and Understanding Urban Environment

Abstract

Understanding urban environments and their spatio-temporal changes is essential for regional and local planning and environmental management. With the emergence of large volumes of various earth observation data, it is important to take advantage of such datasets to improve land cover monitoring, especially the detection for the urbanization process. To reach this goal, a new image processing approach that employs all available historical Landsat images is proposed, and tested in Broome County, NY. Not only can this method support sub-pixel land cover mapping (as in traditional studies), but it also can derive the magnitude, timing and duration of urbanization (which cannot be provided in traditional studies). The Broome County experiment shows that the performance of this processing technique is comparable with, and even slightly better than, the existing data product. The results derived from machine learning methods are also compared, and the land cover information extracted from various satellite images is further utilized in support of urban sustainability studies.

November 15, 2016

2016-11-09T20:21:57-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, November 15, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Ziang Zhang (Electrical and Computer Engineering)
Title: Asymptotical Frequency Synchronization of Kuramoto Oscillators by Topology Evolution

Abstract

From beats generated by cardiac pacemakers to the movement of planetary systems, the idea of synchronization of self-organizing oscillators that are coupled over networks links mathematics with natural phenomenon. Although this general idea was recognized decades ago, whether a particular system can synchronize and how to design controllers to help the synchronization remain unknown in many cases. Since thousands of independent generators across the nation, coupled by transmission lines, oscillate at the same 60Hz frequency, numerous studies suggest that a power system is a perfect example of coupled oscillators. However, several major barriers need to be removed in order to bridge the gap between classical phase coupled oscillators and a realistic power system. This talk will discuss some recent findings from Dr. Zhang’s group.

December 6, 2016

2016-11-29T08:07:18-04:00

Statistical Machine Learning Seminar
Hosted by Department of Mathematical Sciences

Date: Tuesday, December 6, 2016
Time: 12:00-1:00
Room: WH-100E
Speaker: Yang Ning (Cornell University)
Title: A General Framework for High-Dimensional Inference and Multiple Testing

Abstract

We consider the problem of how to control the false scientific discovery rate in high-dimensional models. Towards this goal, we focus on the uncertainty assessment for low dimensional components in high-dimensional models. Specifically, we propose a novel decorrelated likelihood based framework to obtain valid p-values for generic penalized M-estimators. Unlike most existing inferential methods which are tailored for individual models, our method provides a general framework for high-dimensional inference and is applicable to a wide variety of applications, including generalized linear models, graphical models, classifications and survival analysis. The proposed method provides optimal tests and confidence intervals. The extensions to general estimating equations are discussed. Finally, we show that the p-values can be combined to control the false discovery rate in multiple hypothesis testing.