http://www.cs.rutgers.edu/news/colloquia/ DCS Events en-us 1 Copyright (c) 2012 Rutgers, The State University of New Jersey. All rights reserved. http://www.cs.rutgers.edu/news/colloquia/colloquia.php?action=view&colloquium_id=5966 Thu, 09 Feb 2012 19:31:47 -0500 Dr. Jingrui He from IBM Watson Research Center, 02-15-2012 01:30 PM at CBIM 22 (Multipurpose Room) Dr. Jingrui He Tina Eliassi-Rad Colloquium CBIM 22 (Multipurpose Room) Heterogeneity Meets Rarity: Mining Multi-Faceted Diamond 02-15-2012 01:30 PM IBM Watson Research Center Dr. Jingrui He is currently a research staff member at IBM T.J. Watson Research Center. She received her M.Sc and Ph.D degree from Carnegie Mellon University in 2008 and 2010 respectively, both majored in Machine Learning. Her research interests include developing scalable algorithms for heterogeneous learning, rare category analysis, and semi-supervised learning with an emphasis on applications in social media analysis. She is the recipient of IBM Fellowship between 2008 and 2010. She also won the second place in ICDM2010 data mining contest on traffic prediction (both Task 2 and Task 3). She has published over 30 referred articles and served as the organization committee member of ICML, KDD, etc. Many real-world machine learning and data mining problems exhibit both heterogeneity and rarity. Take anomaly detection (e.g., insider threat detection) from various social contexts as an example. While the target abnormal persons may only be a very small portion of the entire population (i.e., rarity), each person can be characterized by rich features, such as hyper-links, texts, social friendship, etc (i.e., feature heterogeneity). Moreover, different types of anomalies, though correlated, may exhibit different statistical characteristics (i.e., task heterogeneity). How can we identify at least one example for a new type of rare category? How can we leverage both feature heterogeneity and task heterogeneity to maximally boost the learning performance? In this talk, I will present our recent work on addressing these two challenges. For the challenge of heterogeneity, I will introduce a graph-based approach for both feature heterogeneity and task heterogeneity. For the challenge of rarity, I will talk about rare category analysis, e.g., how to discover the rare examples with the help of a labeling oracle, how to simultaneously identify both the rare examples and the relevant subspace, etc. DCS http://www.cs.rutgers.edu/news/colloquia/colloquia.php?action=view&colloquium_id=5964 Thu, 09 Feb 2012 19:31:47 -0500 Dr. Hanghang Tong from IBM Watson Research Center, 02-15-2012 12:00 PM at CBIM 22 (Multipurpose Room) Dr. Hanghang Tong Tina Eliassi-Rad Colloquium CBIM 22 (Multipurpose Room) Algorithms to Affect Influence Propagation on Large Graphs 02-15-2012 12:00 PM IBM Watson Research Center Dr. Hanghang Tong is currently a research staff member at IBM T.J. Watson Research Center. Before that, he was a Post-doctoral fellow in Carnegie Mellon University. He received his M.Sc and Ph.D. degree from Carnegie Mellon University in 2008 and 2009, both majored in Machine Learning. His research interest is in large scale data mining for graphs and multimedia. He was a co-PI in NSF sponsored project on virus and influence propagation on large graphs. He is currently a co-PI in two DARPA sponsored projects on computational social science (ADAMS and SMISC). He is an IBM PI and task lead in the Social and Cognitive Networks Academic Research Center (SCNARC) sponsored by Army Research Lab. His current task focuses on composite networks in organization and team performance. He has received several awards, including best research paper award in ICDM 2006 and best paper award in SDM 2008. He has published over 40 referred articles and served as a program committee member of SIGKDD, PKDD, and WWW. Large graphs are everywhere, and they are becoming a prevalent platform for the masses to interact and disseminate a variety of information (e.g., memes, opinions, rumors, etc).Controlling the outcome of such dissemination on a large graph is an interesting problem in many disciplines, such as epidemiology, computer security, marketing, etc. In this talk, we focus on the problem of optimally affecting the outcome of influence propagation by manipulating the underlying graph structure. We show that for a large family of influence propagation models, the problem becomes optimizing the leading eigen-value of an appropriately defined system matrix associated with the underlying graph. We then present two algorithms as the instantiations of such an optimization problem - one to minimize the leading eigen-value (e.g., stopping virus propagation) by deleting nodes from the graph, and the other to maximize the eigen-value (e.g., promoting product adoption) by adding edges to the graph. DCS http://www.cs.rutgers.edu/news/colloquia/colloquia.php?action=view&colloquium_id=5970 Thu, 09 Feb 2012 19:31:47 -0500 Benny Pinkas from Bar Ilan University, Israel, 02-15-2012 11:00 AM at CoRE 431 Benny Pinkas Eric Allender Seminar CoRE 431 Secure Computation on the Web: Computing without Simultaneous Interaction 02-15-2012 11:00 AM Bar Ilan University, Israel Secure computation enables mutually suspicious parties to compute a joint function of their private inputs while providing strong security guarantees. However, its use in practice seems limited. We argue that one of the reasons for this is that the model of computation on the web is not suited to the type of communication patterns needed for secure computation. Specifically, in most web scenarios clients independently connect to servers, interact with them and then leave. This rules out the use of secure computation protocols that require that all participants interact simultaneously. We initiate a study of secure computation in a client-server model where each client connects to the server once and interacts with it, without any other client necessarily being connected at the same time. We point out some inherent limitations in this model and present definitions that capture what can be done. We also present a general feasibility result and several truly practical protocols for a number of functions of interest. All our protocols are based on standard assumptions, and we achieve security both in the semi-honest and malicious adversary models. Joint work with by Shai Halevi and Yehuda Lindell. DCS http://www.cs.rutgers.edu/news/colloquia/colloquia.php?action=view&colloquium_id=5968 Thu, 09 Feb 2012 19:31:47 -0500 Darakhshan J. Mir from Rutgers University, 02-13-2012 01:00 PM at CoRE A (Room 301) Darakhshan J. Mir Rebecca Wright, Tina Eliassi-Rad, Michael Littman and Tomasz Imielinski Qualifying Exam CoRE A (Room 301) Differential Privacy in Various Contexts of Sensitive Data Analysis 02-13-2012 01:00 PM Rutgers University We summarize the deployment of the rigorous notion of differential privacy to enable analysis and use of sensitive data in various contexts. Facilitating the use of sensitive data for research or commercial purposes, in a manner that preserves the privacy of the participating individuals, has always been an important topic in the computer science, statistics, and some social sciences communities. In this article we look at such problems, that arise in various data analyses situations. We apply the notion of differential privacy to provide privacy preserving solutions to these problems while also demonstrating the utility of the privacy-preserving analysis--- in some cases, experimentally and in some theoretically. First, we examine the problem of privately estimating the parameters of a statistical graph model, the stochastic Kronecker graph model, to be able to generate private synthetic graphs that can mimic some statistics of the original graph. Second, we study the problem of differentially private learning and relate it to the PAC-Bayesian learning framework. This enables us to express differentially private learning as a constrained information disclosure problem. Third, we examine privacy in the online streaming scenario using an extension of differential privacy called pan-privacy to propose a useful estimator for the distinct counts statistic, over a stream of data that includes both positive and negative updates. DCS