Rutgers is pleased to host the Yahoo! Seminar Series in Machine Learning.

There is a wide range of research in the area of machine learning at Rutgers and the surrounding area. The goal of this seminar series is to spread awareness of research in this topic inside the university and across disciplines where it is studied and leveraged. Past speakers include Nicholas Belkin, Stephen Jose Hanson, Ying Hung, John Langford, Mike Lesk, Smaranda Muresan, Mor Naaman, Alexander Smola, Umar Syed, and Tong Zhang. The talks from 2009-10 are listed here.

The colloquia is graciously sponsored by Yahoo! and is organized by Tina Eliassi-Rad, Michael Littman, and Michael Pazzani in 2010-11. Meetings are typically held on Wednesdays at 11:00 AM in Room 301 of the CoRE Building and are followed by a pizza lunch with the speaker.

Professors Mor Naaman and Paul Kantor are organizing the 2011-12 seminar series.

Schedule for Academic Year 2010-11

Date Agenda Speaker Audio/Video
September 27, 2010 Learning without Search Geoff Webb
October 19, 2010 What Users Want Alexander Smola
November 1, 2010 Profiling-by-Association: A Resilient Traffic Profiling Solution for Network Trace Data Tina Eliassi-Rad
November 16, 2010 Solving Unsolvable Games Martin Zinkevich
November 30, 2010 Medical Decision-Support System: Optimizing Pattern Recognition of Medical Signal Data W. Art Chaovalitwongse
January 19, 2011 Efficient Bayesian Methods for Hierarchical Clustering Katherine A. Heller
January 26, 2011 Topic Models We Can Believe In: New Approaches to Evaluating Latent Variable Models for Text Analysis David Mimno
February 23, 2011 Machine Learning Algorithms for Real Data Sources, with Applications to Climate Science Claire Monteleoni
March 2, 2011 Learning Feature Hierarchies for Vision Yann LeCun
March 9, 2011 Comprehensive Patient Similarity Learning Jimeng Sun
March 23, 2011 Inferring the Structure and Scale of Modular Networks Jake Hofman
March 30, 2011 Socially Intelligent Machine Learning Haym Hirsh
April 6, 2011 Collective Graph Identification Lise Getoor
April 13, 2011 Some Remarks on the Model Selection Problem Branden Fitelson
April 20, 2011 How ML Relates Victoria Stilwell to George Lucas David L. Roberts
April 27, 2011 Computational Insights into Population Biology Tanya Berger-Wolf
May 4, 2011 Recommender Systems: The Art and Science of Matching Items to Users Deepak Agarwal

Titles, Abstracts, and Bios

  • Wednesday, May 4, 2011 at 11:00 AM, CORE 301

    Recommender Systems: The Art and Science of Matching Items to Users
    Deepak Agarwal, Yahoo! Research

    (slides, audio, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Algorithmically matching items to users in a given context is essential for the success and profitability of large scale recommender systems like content optimization, computational advertising, web search, shopping, and movie recommendation and so on. A key statistical problem that is essential to the success of such systems is to estimate response rates of some rare event (e.g. click-rates, buy rates, etc) when users interact with items. This is a very high dimensional estimation problem since data is obtained by interactions among several heavy-tailed categorical variables. In this talk, I will discuss statistical techniques based on large scale multi-level hierarchical models, some of which have been deployed and are successfully recommending articles and ads to users on Yahoo! websites. The methods described are based reduced rank logistic regression, probabilistic matrix factorization, supervised Latent Dirichlet Allocation, and multi-hierarchy smoothing.

    Bio: Deepak Agarwal is a statistician at Yahoo! who is interested in developing statistical and machine learning methods to enhance the performance of large scale recommender systems. Deepak and his collaborators significantly improved article recommendation on several Yahoo! websites, most notably on the Yahoo! front page. He also works closely with teams in computational advertising, yet another large scale recommender system. He serves as associate editor for the Journal of American Statistical Association and has received four best paper awards in the past.

  • Wednesday, April 27, 2011 at 11:00 AM, CORE 301

    Computational Insights into Population Biology
    Tanya Berger-Wolf, University of Illinois, Chicago

    (slides, audio, audio/slides giant, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Computation has fundamentally changed the way we study nature. Recent breakthroughs in data collection technology, such as GPS and other mobile sensors, high definition cameras, GIS, and genotyping, are giving biologists access to data about wild populations, from genetic to social interactions, which are orders of magnitude richer than any previously collected. Such data offer the promise of answering some of the big questions in population biology: Why do animals form social groups and how do genetic ties affect this process? Which individuals are leaders and to what degree do they control the behavior of others? How do social interactions affect the survival of a species? Unfortunately, in this domain, our ability to analyze data lags substantially behind our ability to collect it. In this talk I will show how computational approaches can be part of every stage of the scientific process, from data collection (identifying individual zebras from photographs) to hypothesis formulation (by designing a novel computational framework for analysis of dynamic social networks). I will also show lots of pictures of exotic animals.

    Bio: Dr. Tanya Berger-Wolf is an Associate Professor in the Department of Computer Science at the University of Illinois at Chicago where she heads the Computational Population Biology Lab. Her research interests are in applications of computational techniques to problems in population biology of plants, animals, and humans, from genetics to social interactions. Dr. Berger-Wolf received her Ph.D. in Computer Science from University of Illinois at Urbana-Champaign in 2002. She spent two years as a postdoctoral fellow at the University of New Mexico working on computational phylogenetics and a year at the Center for Discrete Mathematics and Theoretical Computer Science (DIMACS) doing research in computational epidemiology. She has received numerous awards for her research and mentoring, including the US National Science Foundation CAREER Award in 2008 and the UIC Mentor of the Year Award in 2009.

  • Wednesday, April 20, 2011 at 11:00 AM, CORE 301

    How ML Relates Victoria Stilwell to George Lucas
    David L. Roberts, North Carolina State University

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: The "Computers As Social Actors" studies have provided significant evidence that humans apply social norms to their interactions with technology. As a result, the richness of the interactions we expect with computers has grown significantly over the years, creating increasingly complex challenges for designers. Often, these designers are subject matter experts, rather than highly-trained technical experts, which can further complicate these challenges. Research in AI and ML has sought to create techniques that can work in concert with designers to bring about meaningful interactive experiences. The end result is the "co-creation" of experiences that arises from a partnership among the designer, the computational artifact, and the experience participant.

    In this talk we will examine work in AI and ML that is inspired by social and behavioral psychology and that leverages the psychological aspects of interactions with technology to co-create desirable experiences. We will discuss two application areas that will establish a subtle link between the work of two seemingly disparate people: Victoria Stilwell, a renowned dog trainer, and George Lucas, a renowned film and computer game producer. Each, accomplished in their own right, works in a completely different realm. But, when we dig down under the hood, we will find that machine learning is a lens through which their work can be viewed as highly similar.

    Bio: David Roberts is an Assistant Professor of Computer Science at North Carolina State University. He received his Ph.D. from the College of Computing at Georgia Tech in 2010. His research interests lie at the interface of artificial intelligence, machine learning, social/behavioral psychology, and human-computer interaction with an application focus on interactive virtual experiences.

  • Wednesday, April 13, 2011 at 11:00 AM, CORE 301

    Some Remarks on the Model Selection Problem
    Branden Fitelson, Rutgers University

    (slides, audio, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Model selection has received a lot of attention in the statistical/computer science literature the past 40 years or so. This talk will provide an overview of "the" model selection problem, and various paradigms that have been developed for understanding (and attacking) it.

    Bio: Branden Fitelson is currently an Associate Professor of Philosophy at Rutgers University. He is also a member of the Rutgers Center for Cognitive Science (RuCCS). Prior to joining Rutgers, Branden was an Associate Professor of Philosophy at University of California, Berkeley. Branden has been a co-organizer of the annual Formal Epistemology Workshops (FEW), since 2003. He also serves as a subject editor for the Formal Epistemology section of the Stanford Encyclopedia of Philosophy, and he is one of the editors of the Journal of Philosophical Logic. His recent research is at the intersection of philosophy of science, epistemology, logic, and cognitive science (viz., formal epistemology). He has published over 50 articles and has edited several special issues of journals on these topics. Branden received his PhD from University of Wisconsin-Madison.

  • Wednesday, April 6, 2011 at 11:00 AM, CORE 301

    Collective Graph Identification
    Lise Getoor, University of Maryland, College Park

    (slides, audio, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: The importance of network analysis is growing across many domains, and is fundamental in understanding online social interactions, biological processes, communication, ecological, financial, transportation networks, and more. In most of these domains, the networks of interest are not directly observed, but must be inferred from noisy and incomplete data, data that was often generated for purposes other than scientific analysis. In this talk, I will describe graph identification, the process of inferring the hidden network from noisy observational data. In particular, I will describe a collective approach to graph identification, which interleaves the necessary steps in the accurate reconstruction of the network.

    Joint work with Galileo Namata and Stanley Kok, University of Maryland.

    Bio: Lise Getoor is an Associate Professor at the University of Maryland, College Park. Her main research areas are machine learning and reasoning under uncertainty and she has also done work in areas such as database management, social network analysis and visual analytics. She is PC co-chair of ICML 2011, and has served as senior PC or PC member for conferences including AAAI, ICML, IJCAI, ICWSM, KDD, SIGMOD, UAI, VLDB, and WWW. She is an ACM Transactions on Knowledge Discovery and Data Associate Editor, was a Journal of Artificial Intelligence Research Associate Editor, and Machine Learning Journal Action Editor. She is on the board of the International Machine Learning Society, and has served on the AAAI Council. She is a recipient of an NSF Career Award, was a Microsoft New Faculty Fellow finalist and was awarded a National Physical Sciences Consortium Fellowship. She received her PhD from Stanford University, her Masters degree from University of California, Berkeley, and her BS from the University of California, Santa Barbara.

  • Wednesday, March 30, 2011 at 11:00 AM, CORE 301

    Socially Intelligent Machine Learning
    Haym Hirsch, Rutgers University

    (audio, audio/slides big, audio/slides small)

    Faculty Host: Prof. Michael Littman

    Abstract: The pervasive success of computing has given rise to new questions at the intersection of computer and behavioral sciences . I use the term "socially intelligent computing" to refer to two related questions in this area. First, how can we create systems that interact with people in ways that appear "socially intelligent"? Second, how can interactions among people, and between peopl e and increasingly sophisticated computing, at scales ranging from a single person and machine to Internet-scale systems of people a nd computing, give rise to new forms of "intelligent computing"? In this talk I'll give an overview of key ideas and projects by my self and others in socially intelligent computing, with a particular focus on examples at the intersection of machine learning and s ocially intelligent computing -- both of machine learning enabling new forms of socially intelligent computing, and of socially inte lligent computing facilitating new forms of machine learning.

    Bio: Haym Hirsh is Professor of Computer Science at Rutgers University and a Visiting Scholar at the Sloan School of Management and Center for Collective Intelligence at MIT. His research focuses on foundations and applications of machine learning, data mining, a nd information retrieval. Haym received his BS from the Mathematics and Computer Science Departments at UCLA and his MS and PhD fro m the Computer Science Department at Stanford University. He has held visiting positions at Bar-Ilan University, CMU, MIT, NYU, and the University of Zurich; from 2006-2010 he served as Director of the Division of Information and Intelligent Systems at the Nation al Science Foundation.

  • Wednesday, March 23, 2011 at 11:00 AM, CoRE 301

    Inferring the Structure and Scale of Modular Networks
    Jake Hofman, Yahoo! Research

    (audio, audio/slides giant, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Much recent work has focused on community detection, or the task of identifying sets of similar nodes from network topology. Underlying this work is the implicit assumption that inferred communities inform node attributes or function in a meaningful and useful sense. We investigate these ideas by phrasing community detection as Bayesian inference, which gives rise to a scalable and efficient variational algorithm for fitting and comparing network models. We show how several existing methods for community detection can be described as variant, special, or limiting cases of our work, and how the method overcomes the "resolution limit" problem, accurately recovering both the structure and scale of modular networks. We apply the resulting algorithm to several real networks and study the relationship between identified topological communities and known node attributes.

    Bio: Jake Hofman is a Research Scientist and member of the Human Social Dynamics Group at Yahoo! Labs New York. His work involves data-driv en modeling of complex systems, focusing on applications of machine learning and statistical inference to large-scale network data. He hold s a B.S. in Electrical Engineering from Boston University and a Ph.D. in Physics from Columbia University.

  • Wednesday, March 9, 2011 at 11:00 AM, CoRE 301

    Comprehensive Patient Similarity Learning
    Jimeng Sun, IBM TJ Watson

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: The objective of patient similarity is to quantitatively measure how similar patients are to each other. The challenges of comprehensive patient similarity are the following:

  • How to integrate multiple sources of clinical information for patient similarity computation?
  • How to compare patients at different stages of disease progression?
  • How to leverage expert feedback into the similarity computation?
  • How to incrementally update the existing patient similarity functions as new data arrive?
  • How to present the similarity in an intuitive way?
  • In this work, we will present the comprehensive patient similarity framework that answers those questions. The core of the framework is the combination of advanced analytic algorithms and novel visualization techniques. We also present some empirical studies on real patient data from a large healthcare network over 200K patients. Finally, we envision the patient similarity framework can enable many important clinical applications such as comparative effectiveness research (CER), treatment recommendation, and physician comparison model.

    Bio: Jimeng Sun is a Research Staff Member at IBM TJ Watson. He received the MS and PhD degrees in Computer Science from Carnegie Mellon University in 2006 and 2007. His research interests include data mining for health care applications, medical informatics, social network analysis, visual analytics, and data streams. He has received the best research paper award at ICDM 2008, the KDD 2007 dissertation award (runner-up), the best research paper award at SDM 2007. He has published over 40 refereed articles and two book chapters. He filed eight patents and has given four tutorials. He has served as the program committee member of SIGKDD, ICDM, SDM and CIKM and a reviewer for AMIA, TKDE, VLDB, and ICDE. He has co-chaired the workshops on large-scale data mining: theory and applications in KDD 2010 and ICDM 2009, the workshop on large-scale Analytics for Complex Instrumented Systems on ICDM 2010, and the workshop on Visual Analytics in Health Care in VisWeek 2010. He also co-edited the journal special

  • Wednesday, March 2, 2011 at 11:00 AM, CoRE Auditorium (Room 101)

    Learning Feature Hierarchies for Vision
    Yann LeCun, New York University

    (slides, audio, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Intelligent perceptual tasks such as vision and audition require the construction of good internal representations. Theoretical and empirical evidence suggest that the perceptual world is best represented by a multi-stage hierarchy in which features in successive stages are increasingly global, invariant, and abstract. An important challenge for Machine Learning is to devise "deep learning" methods than can automatically learn good feature hierarchies from labeled and unlabeled data. A class of such methods that combine unsupervised sparse coding, and supervised refinement will be described. We demonstrate the use of deep learning methods to train convolutional networks (ConvNets). ConvNets are biologically-inspired architectures consisting of multiple stages of filter banks, non-linear operations, and spatial pooling operations, analogous to the simple cells and complex cells in the mammalian visual cortex. A number of applications will be shown through videos and live demos, including a category-level object recognition system that can be trained on the fly, a pedestrian detector, and system that recognizes human activities in videos, and a trainable vision system for off-road mobile robot navigation. A new kind of "dataflow" computer architecture, dubbed NeuFlow, was designed to run these algorithms (and other vision and recognition algorithms) in real time on small, embeddable platforms. an FPGA implementation of NeuFlow running various vision applications will be demonstrated. An ASIC is being designed in collaboration with e-lab at Yale, which will be capable of 2000 Giga-operations per second for less than 5 watts.

    Bio: Yann LeCun is Silver Professor of Computer Science and Neural Science at the Courant Institute of Mathematical Sciences and at the Center for Neural Science of New York University. He received an Electrical Engineer Diploma from Ecole Supérieure d'Ingénieurs en Electrotechnique et Electronique (ESIEE), Paris in 1983, and a PhD in Computer Science from Université Pierre et Marie Curie (Paris) in 1987. After a postdoc at the University of Toronto, he joined AT&T Bell Laboratories in Holmdel, NJ, in 1988. He became head of the Image Processing Research Department at AT&T Labs-Research in 1996, and joined NYU in 2003, after a brief period as Fellow at the NEC Research Institute in Princeton. His current interests include machine learning, computer vision, pattern recognition, mobile robotics, and computational neuroscience. He has published over 150 technical papers on these topics as well as on neural networks, handwriting recognition, image processing and compression, and VLSI design. His handwriting recognition technology is used by several banks around the world to read checks. His image compression technology, called DjVu, is used by hundreds of web sites and publishers and millions of users to distribute and access scanned documents on the Web, and his image recognition technique, called Convolutional Network, has been deployed by companies such as Google, Microsoft, NEC, France Telecom and several startup companies for document recognition, human-computer interaction, image indexing, and video analytics. He has been on the editorial board of IJCV, IEEE PAMI, IEEE Trans on Neural Networks, was program chair of CVPR'06, and is chair of the annual Learning Workshop. He is on the science advisory board of Institute for Pure and Applied Mathematics, and is the co-founder of MuseAmi, a music technology company.

  • Wednesday, February 23, 2011 at 11:00 AM, CoRE 301

    Machine Learning Algorithms for Real Data Sources, with Applications to Climate Science
    Claire Monteleoni, Columbia University

    (slides, audio, audio/slides big, audio/slides small)

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Profuse amounts of digital data are being generated from a myriad of sources. Satellites record vast sequences of high-dimensional images, and environmental sensors continually measure temperature and atmospheric gases. The growing reliance on the Internet for daily tasks also fuels this rapid data growth. Real data sources increasingly pose interesting and urgent challenges for machine learning algorithm design; the data can be vast, high-dimensional, streaming, noisy, time-varying, private, or it may combine these and other attributes.

    In this talk, I will discuss my work on designing machine learning algorithms, with formal performance guarantees, motivated by fundamental properties of real data sources. Addressing the problems of learning from data streams, learning from raw (unlabeled) data, and learning from private data, I will survey my work on online learning, active learning, clustering, and privacy-preserving machine learning. I will also motivate the study of Climate Informatics: accelerating discovery in Climate Science with machine learning. I will present an algorithm for online learning with expert predictors, and demonstrate its performance in tracking climate models, a new application in Climate Informatics.

    Bio: Dr. Claire Monteleoni is research faculty in the Center for Computational Learning Systems, and adjunct faculty in the Department of Computer Science, at Columbia University in the City of New York. Prior to joining Columbia, she was a postdoc in Computer Science and Engineering, at the University of California, San Diego. She completed her PhD in 2006 and her Masters in 2003, in Computer Science, at MIT. She did her undergraduate work, in Earth and Planetary Sciences, at Harvard University. Her research focus is on Machine Learning theory and algorithms, in particular: Learning from Data Streams, Clustering, Active Learning, and Privacy-Preserving Machine Learning. She has recently started working on Climate Informatics: accelerating discovery in Climate Science with Machine Learning. Her work in this area has received a Best Application Paper Award, and has been presented at an Expert Meeting of the Intergovernmental Panel on Climate Change (IPCC), a panel formed by the UN, that shared the 2007 Nobel Peace Prize.

  • Wednesday, January 26, 2011 at 11:00 AM, CoRE 301

    Topic Models We Can Believe In: New Approaches to Evaluating Latent Variable Models for Text Analysis
    David Mimno, Princeton University

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Models that identify latent semantic dimensions in text, such as statistical topic models, are a popular method for approaching today's massive document collections. Much of the research in computer science, however, has not evaluated how real users respond to such models. In this talk I will discuss how my research in topic modeling has been influenced by interactions with users outside computer science, how those reactions can be translated into evaluation metrics, and how those evaluations have guided further model development. In particular I will discuss the effect of prior distributions on Bayesian mixed membership models, automated semantic evaluations of topic distributions, and model fit testing using posterior predictive checks.

    Bio: David Mimno is a postdoctoral researcher in the department of Computer Science at Princeton University. He holds a bachelors degree in Classics and Computer Science from Swarthmore College, and completed his PhD at UMass, Amherst in 2010. Prior to graduate school, he worked for the Perseus Digital Library at Tufts University. His work has been supported by the Mellon Foundation, the National Institutes of Health, and a Google Digital Humanities Research award.

  • Wednesday, January 19, 2011 at 11:00 AM, CoRE 301

    Efficient Bayesian Methods for Hierarchical Clustering
    Katherine A. Heller, University of Cambridge & University of Pennsylvania

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: I will present Bayesian Hierarchical Clustering (BHC), an efficient agglomerative hierarchical clustering method based on evaluating marginal likelihoods of a probabilistic model. BHC has several advantages over traditional distance-based agglomerative clustering algorithms. (1) It defines a probabilistic model of the data which can be used to compute the predictive distribution of a test point and the probability of it belonging to any of the existing clusters in the tree. (2) It uses a model-based criterion to decide on merging clusters rather than a global distance metric. (3) Bayesian hypothesis testing is used to decide which merges are advantageous and to output the recommended depth of the tree. (4) BHC can be interpreted as a fast bottom up approximate inference method for a Dirichlet process mixture model (DPM). It approximates the marginal likelihood of a DPM by summing over exponentially many clusterings of the data in polynomial time.

    I will also present a new generalization of BHC which discovers rose trees. One of the drawbacks of BHC, and hierarchical clustering methods which are restricted to binary branching structure in general, is that they necessarily discover hierarchy structure which may not be present in the data, thereby resulting in needless cascades. Our Bayesian Rose Trees (BRT) method alleviates this problem by efficiently discovering hierarchies with arbitrary branching structure at each node.

    This is joint work with Zoubin Ghahramani, Yang Xu, Charles Blundell, and Yee Whye Teh.

    Bio: Katherine A. Heller is an EPSRC Postdoctoral Research Fellow in the Engineering Department at the University of Cambridge. She received a B.S. degree in Computer Science and Applied Mathematics and Statistics from SUNY Stony Brook, an M.S. degree in Computer Science from Columbia University, and a Ph.D. in Machine Learning from the Gatsby Unit, University College London. Her research interests lie in machine learning, Bayesian statistics, nonparametric Bayesian methods, and clustering, with applications to cognitive science and information retrieval.

  • Tuesday, November 30, 2010 at 11:00 AM, CoRE 301

    Medical Decision-Support System: Optimizing Pattern Recognition of Medical Signal Data
    W. Art Chaovalitwongse, Rutgers University

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: I will discuss some recent advances in optimization and data mining used to develop a new pattern recognition framework. This work relates to medical data signal processing apparatus and computational framework, where optimization and data mining techniques are employed to analyze medical signal data as advanced medical decision-support systems. The ultimate goal of this research is to improve the current medical diagnosis and prognosis by assisting the physicians in recognizing (data-mining) abnormality patterns in medical data. The diagnosis of epilepsy and brain disorders is a case point in this study. Brain diagnoses largely deal with neurophysiological signals such as electroencephalograms (EEGs), in which the brain's functional properties are encrypted in a form of large-scale multichannel time series. The proposed framework will be used to predict epileptic seizures. Specifically, we will employ our framework to recognize seizure precursors and identify seizure susceptibility (pre-seizure) periods.

    Acknowledgements: This work is partially supported by NSF under CAREER grant (0546574) and Rutgers Computing Coordination Council (CCC).

    Bio: Dr. Wanpracha Art Chaovalitwongse is an Associate Professor of Industrial and Systems Engineering at Rutgers University where he has been on the faculty since 2005. He received his Bachelor's degree in Telecommunication Engineering from King Mongkut Institute of Technology at Ladkrabang, Thailand, in 1999 and M.S. and Ph.D. degrees in Industrial and Systems Engineering from University of Florida in 2000 and 2003. Before joining Rutgers, he worked at the Corporate Strategic Research, ExxonMobil Research & Engineering, where he managed research in developing efficient mathematical models and novel statistical data analyses for upstream oil exploration and downstream business operations in multi-continent oil transportation. His lab conducts basic science, applied, and translational research at the interface of engineering, medicine, and other emerging disciplines. He holds three patents on novel optimization techniques adopted in the development of seizure prediction system. His academic honors include 2003 Excellence in Research from the University of Florida, 2006 NSF CAREER Award, 2007 Notable Alumni of King Mongut's Institute Technology at Ladkrabang, 2004 & 2008 (2-times winner) William Pierskalla Best Paper Award by the Institute for Operations Research and the Management Sciences (INFORMS), 2009 Outstanding Service Award by the Association of Thai Professionals in America and Canada, and 2010 Rutgers Presidential Fellowship for Teaching Excellence.

  • Tuesday, November 16, 2010 at 11:00 AM, CoRE 301

    Title: Solving Unsolvable Games
    Martin Zinkevich, Yahoo! Research

    Faculty Host: Prof. Michael Littman

    Abstract: I will discuss the 2010 Lemonade Stand Competition, where nine teams competed in a three-player symmetric, constant sum, unsolvable game. A game is unsolvable if there is an equilibrium A for Alice, an equilibrium B for Belle, and an equilibrium C for Carly, where if every player plays her part of her equilibrium, then the resulting combined play is not in equilibrium. Moreover, in this game, the "safe" strategy guarantees a very small fraction of the total utility (the team that played the safe strategy came in last). On the other hand, simple strategies, such as choosing a fixed strategy for a few rounds, and then moving randomly if the utility obtained is not above the average, came in near the top. These simple strategies benefit from some of the more sophisticated strategies that learned to cooperate. This competition has been successful in its short-term goal of building a publicly available library of intelligent agents. The long-term goal of this competition is to help scientists develop a new language of describing intelligent behavior, that can be used in understanding human experiments, designing theoretical models of intelligence, and building more general intelligent agents in the future.

    Bio: Martin Zinkevich is a Senior Research Scientist at Yahoo! Research. He received his Ph. D. in computer science at Carnegie Mellon in 2004. He works on both applied and theoretical large scale machine learning and game theory applied to advertising, antispam, and poker.

  • Monday, November 1, 2010 at 2:00 PM, CBIM 22

    Profiling-by-Association: A Resilient Traffic Profiling Solution for Network Trace Data
    Tina Eliassi-Rad, Rutgers University

    Faculty Host: Prof. Michael Pazzani

    Abstract: Profiling network traffic is becoming an increasingly hard problem since users and applications are avoiding detection using traffic obfuscation and encryption. The key question addressed here is: Is it possible to profile network traffic without relying on its packet and flow level information, which can be obfuscated? We propose a novel approach, called Profiling-By-Association (PBA), that uses only the IP-to-IP communication graph and information about some applications used by few IP-hosts (a.k.a. seeds). The key insight is that IP-hosts tend to communicate more frequently with hosts involved in the same application forming communities (or clusters). Profiling few members within a cluster can “give away” the whole community. Following our approach, we develop different algorithms to profile network traffic and evaluate them on real-traces from four large networks. We show that PBA’s accuracy is on average around 90% with knowledge of only 1% of all the hosts in a given data set and its runtime is on the order of minutes (~ 5).

    This is joint work with Marios Iliofotou, Guowu Xie, and Michalis Faloutsos at UC Riverside and Brian Gallagher at LLNL.

    Bio: Tina Eliassi-Rad is an Assistant Professor at the Department of Computer Science at Rutgers University. She is also a member of the Rutgers Center for Computational Biomedicine, Imaging, and Modeling (CBIM) and Rutgers Center for Cognitive Science (RuCCS). Until September 2010, Tina was a Member of Technical Staff at Lawrence Livermore National Laboratory. Tina earned her Ph.D. in Computer Sciences (with a minor in Mathematical Statistics) at the University of Wisconsin-Madison in 2001. Broadly speaking, Tina's research interests include machine learning, data mining, and artificial intelligence. Her work has been applied to the World-Wide Web, text corpora, large-scale scientific simulation data, and complex networks. Tina is an action editor for the Data Mining and Knowledge Discovery Journal. She received a US DOE Office of Science Outstanding Mentor Award in 2010.

  • Tuesday, October 19, 2010 at 11:00 AM, CoRE 301

    What Users Want
    Alexander Smola, Yahoo! Research

    Faculty Host: Prof. Tong Zhang

    Abstract: User modeling is a key component in understanding how humans interact with internet properties. In this talk I present a set of nonparametric Bayesian modeling techniques to address these issues. In particular I will cover problems of large scale inference in Latent Dirichlet Allocation, the integration of information between several sources of data, and time-dependence in modeling user activity. The talk will touch both on the statistical modeling issues involved and on the engineering problems on the systems side to make such algorithms work at internet scale.

  • Monday, September 27, 2010 at 2:00 PM, CBIM 22

    Learning without Search
    Geoff Webb, Monash University

    Faculty Host: Prof. Tina Eliassi-Rad

    Abstract: Machine learning is classically conceived as search through a hypothesis space for a hypothesis that best fits the training data. In contrast, naive Bayes performs no search, extrapolating an estimate of a high-order conditional probability by composition from lower-order conditional probabilities. In this talk I show how this searchless approach can be generalised, creating a family of learners that provide a principled method for controlling the bias/variance trade-off. At one extreme very low variance can be achieved as appropriate for small data. Bias can be decreased with larger data in a manner that ensure Bayes optimal asymptotic error. These algorithms have the desirable properties of
       - training time that is linear with respect to training set size,
       - supporting parallel and anytime classification,
       - allowing incremental learning,
       - providing direct prediction of class probabilities,
       - supporting direct handling of missing values, and
       - robust handling of noise.
    Despite being generative, they deliver classification accuracy competitive with state-of-the-art discriminative techniques.

    Bio: Geoff Webb is a Professor of Information Technology Research at Monash University, where he heads the Centre for Research in Intelligent Systems. Prior to Monash he held appointments at Griffith University and then Deakin University, where he received a personal chair. His primary research areas are machine learning, data mining, and user modelling. His commercial data mining software, Magnum Opus, incorporates many techniques from his association discovery research. Many of his learning algorithms are included in the widely-used Weka machine learning workbench. He is editor-in-chief of the highest impact data mining journal, Data Mining and Knowledge Discovery, co-editor of the Encyclopedia of Machine Learning (to be published by Springer), a member of the advisory board of Statistical Analysis and Data Mining and a member of the editorial boards of Machine Learning and ACM Transactions on Knowledge Discovery in Data.