Interface '01 Sessions

Session Organizer: Leland Wilkinson
Session Time: Thursday, June 14th, 10:45am - 12:30pm
Room Location: Santa Ana Room
Session Chair: Leland Wilkinson

Making Trees Interactive: KLIMT

Author:: Simon Urbanek (University of Augsburg, Germany)
Abstract:: Trees are a valuable way of displaying structure in data sets. Adding interactive tools would make them even more valuable. This paper describes our prototype software, KLIMT (Classification - Interactive Methods for Trees), for interactive graphical analysis of trees. The research is work in progress and there are many different possible options. What features do analysts want?

Some Graphics for Recursive Partitioning

Author:: Daniel B. Carr (George Mason University)
Abstract:: This paper presents some graphical templates for constructing, describing, and studying recursive partitioning trees. Some of the templates have multiple applications. For example, the layout approaches for showing random trees are generally applicable to the display of clusters. Thus the paper may be of interest to those who make little use of recursive partitioning.

Flexible Models for Prediction

Session Organizer: Jerome Friedman and Padhraic Smyth
Session Time: Thursday, June 14th, 10:45am - 12:30pm
Room Location: Costa Mesa Room
Session Chair: David Madigan, Rutgers

Predictive Data Mining with Multiple Additive Regression Trees

Author:: Jerome H. Friedman (Stanford University)
Abstract:: Predicting future outcomes based on past observational data is a common application in data mining. The primary goal is usually predictive accuracy, with secondary goals being speed, ease of use, and interpretability of the resulting predictive model. New automated procedures for predictive data mining, based on ``boosting'' (CART) regression trees, are described. The goal is a class of fast ``off-the-shelf'' procedures for regression and classification that are competitive in accuracy with more customized approaches, while being fairly automatic to use (little tuning), and highly robust especially when applied to less than clean data. Tools are presented for interpreting and visualizing these multiple additive regression tree (MART) models.

Towards Understanding Boosting

Author:: Bin Yu (UC Berkeley)
Abstract:: Boosting is a very effective computational procedure to improve upon an initial estimator/classifier (weak learner). It comes from machine learning and has had impressive successes on large real data sets.This talk takes the recent gradient descent point of view of boosting to understand L2Boost in regression and classification, especially its overfitting resistance. In particular, we derive L2Boost's interesting exponential bias and variance trade-off in regression, and approximate the 0-1 generalization error by its smoothed version to argue for the importance of the bias-reduction in classification, for the overfitting-resistance of the 0-1 loss function through tapering, and for the failure of attempts to decompose the 0-1 loss as a sum of bias and variance. Moreover, we put forward the thesis that the (proper) L2Boost in classification is not worse, or maybe even better, than LogitBoost or AdaBoost. We conclude that the exponentially-diminishing variance (and centered higher moments) plus the overfitting-resistance of the 0-1 loss are responsible for the overfitting resistance of L2Boosting (and possibly other forms of boosting) in classification.

Why Does Model Averaging Work?

Author:: Yoav Freund (AT\&T Labs)
Abstract:: The last few years have seen an increased interest in model averaging techniques such as bagging and boosting. To the statistician, the most interesting aspect of these techniques is their resistance to over-fitting. I will describe two theoretical explanations to this phenomenon. One explanation applies to boosting (and other margin-based methods such as SVM). The other applies to bagging (and other pseudo-Bayesian techniques).

Model-Based Clustering

Session Organizer: Adrian Raftery
Session Time: Thursday, June 14th, 10:45am - 12:30pm
Room Location: Viejo Room
Session Chair: Adrian Raftery, University of Washington

A Model Based Approach to Text Categorization and Clustering

Author:: Alejandro Murua (University of Washington and Insightful Corporation)
Abstract:: In this work we develop a complete methodology for document classification and clustering. Documents can be represented by high-dimensional vectors of term/word frequencies. We study in depth the effect of dimensionality reduction (via Principal Components Analysis), term weighting and frequency transformation on both classification and clustering tasks. We conclude that increasing the feature space dimension beyond a certain critical value depending on the complexity and size of the data does not improve performance, but worsens it; and applying a logarithm or square-root transformation to the term frequencies reduces error rates. We used these findings to construct a model based document clustering (MBDC) algorithm. This explicitly models the data as being drawn from a Gaussian mixture. This is used to construct clusters based on the likelihood of the data, and to classify documents according to the Bayes rule. One main advantage of our approach is the ability to automatically select the number of clusters present in the document collection via Bayes factors. Our experiments with the Topic Detection and Tracking Corpus demonstrates the ability of MBDC to choose a sensible number of clusters as well as meaningful partitions of the data. Moreover, motivated by the document clustering problem, we introduce a novel algorithm to cluster high dimensional large data sets. This algorithm uses model based clustering in the context of splitting the data into more manageable subsets by way of fractionation. An extension to this method (model based re-fractionation) is also proposed.

Model-Based Clustering for Gene Expression Data

Author:: Ka Yee Yeung (University of Washington)
Abstract:: Many biologists are excited about emerging DNA microarray technologies, since they make it possible for the first time to study the simultaneous variations in activities of thousands of genes. There is a great need to develop analytical methodologies to extract the information contained in these rapidly growing data sets. Because of the large number of genes and the complexity of biological networks, clustering is a useful exploratory technique for analysis of such data. Among the many clustering algorithms that have been proposed, model-based clustering stands out as one of the few with rigorous probabilistic underpinnings. Namely, it assumes that the data is generated by a mixture of underlying probability distributions such as multivariate normal distributions. This Gaussian mixture model has been shown to be a powerful tool for many applications. We will present our experiences in applying model-based clustering algorithms to gene expression data, compare it to some of the many other clustering approaches that have been tried, and discuss challenges remaining in this field.

Model-Based Clustering in Multidimensional Scaling

Author:: Man-Suk Oh (Ewha Women's University)
Abstract:: Multidimensional scaling is widely used to handle data which consists of similarity or dissimilarity measures between pairs of objects. Major problems in multidimensional scaling are object configuration, choice of dimension, and clustering of objects. We propose a Bayesian approach to deal with these problems, using a mixture of multivariate normal distributions as a prior distribution of the object coordinates. A Markov chain Monte Carlo algorithm is used to estimate the object configuration and group membership simultaneously. A simple criterion to determine the dimension of object configuration and the number of groups is proposed.

Time-Series and Sequential Analysis (Contributed Session)

Session Time: Thursday, June 14th, 10:45am - 12:30pm
Room Location: Capistrano Room
Session Chair: Laszlo Engelman, UC Irvine

Bayesian Inference in a New Class of Multi-Scale Time Series Models

Author:: Marco A. R. Ferreira (Duke University)
Abstract:: We introduce a class of multi-scale models for time series. The novel framework couples ``simple'' standard Markov models for the time series stochastic process at different levels of aggregation, and links them via ``error'' models to induce a new and rich class of structured linear models reconciling modelling and information at different levels of resolution. Jeffrey's rule of conditioning is used to revise the implied distributions and ensure that the probability distributions at different levels are strictly compatible. Our construction has several interesting characteristics: a variety of autocorrelation functions resulting from just a few parameters, the ability to combine information from different scales, and the capacity to emulate long memory processes. There are at least three uses for our multi-scale framework: to integrate the information from data observed at different scales; to induce a particular process when the data is observed only at the finest scale; as a prior for an underlying multi-scale process. Bayesian estimation based in MCMC analysis is developed, and issues of forecasting are discussed. Two interesting applications are presented. In the first application, we illustrate some basic concepts of our multiscale class of models through the analysis of the flow of a river. In the second application we use our multiscale framework to model daily and monthly log-volatilities of exchange rates.

A Variable Memory Markovian Modeling Approach to Unsupervised Sequence Segmentation

Author:: Gill Bejerano (Hebrew University)
Abstract:: We outline a novel unsupervised sequence segmentation algorithm motivated by information theoretic principles. The algorithm, which segments the sequences into alternating Variable Memory Markov sources, is based on competitive learning between the Markov models, implemented as Prediction Suffix Trees using the minimum description length principle. By applying a model clustering procedure, based on rate distortion theory combined with deterministic annealing, we obtain a hierarchical segmentation of sequences between alternating Markov sources. The algorithm seems to be self regulated and automatically avoids over segmentation. The method is applied successfully to unsupervised segmentation of texts into languages where it is able to infer both the number of languages and the language switching points. When applied to protein sequence families, we demonstrate the method's ability to identify biologically meaningful sub-sequences within the proteins, which correspond to important functional sub-units, known as protein domains.

Causal Investigation of Time Series Using Graphical Modelling

Author:: Marco Reale (University of Canterbury)
Abstract:: (No abstract available)

Evaluating Sequential Tests for A Class of Stochastic Processes

Author:: Xiaoping Xiong (St. Jude Children's Research Hospital)
Abstract:: We propose computation methods for sequential tests for a class of stochastic processes for which the probability density of test statistic can be factorized into a product of a known likelihood function that is independent of stopping rule and a conditional probability that is independent of parameters. The proposed methods improve accuracy and efficiency of computation and enable us to valuate properties of special interests in sequential tests such as the probability of discordance between a sequential test and the nonsequential test at the last stage of the sequential test. We give examples for evaluating sequential tests on information time with normal outcomes.

A Comparison of Reversible Jump MCMC Algorithms for DNA Sequence Segmentation Using Hidden Markov Models

Author:: R. J. Boys (Newcastle University)
Abstract:: Most DNA sequences display evidence of compositional heterogeneity in the form of patches or domains of similar structure. In this paper we address the problem of identifying such regions of homogeneous composition in genome sequences by using hidden Markov models.

Office of Naval Research Overview

Session Organizer: Wendy Martinez
Session Time: Thursday, June 14th, 10:45am - 12:30pm
Room Location: Laguna Room
Session Chair: Wendy Martinez, ONR

Functional Analysis of Computer Network Data

Author:: Jeff Solka (NSWCDD)
Abstract:: This talk we focus on some of our recent efforts in the application of functional data analysis methods to mail and web server access data. The application of cluster analysis to the data will be illustrated. Methods for the visualization of the data will also be presented. If time permits results that illustrate the identification of outliers within the data set will also be given.

Inferring Internal Losses and Delays in Communication Networks from Edge Measurements

Author:: Robert Nowak (Rice University)
Abstract:: Optimizing communication network performance and detecting attacks and intrusions requires knowledge of loss rates and queueing delays at different points in the network. However, it is impractical to directly monitor packet losses and delays at each and every router. Measurements at the edge of the network (sources and receivers) are relatively easy and inexpensive in comparison. Consequently, it is natural to consider the following inverse problem: from edge-based measurements can we infer the loss rates and delays experienced at internal points in the network ? This paper presents a unified formulation of the problems of internal loss and delay estimation, and describes an expectation-maximization algorithm for computing maximum likelihood estimates. We also propose a new method for jointly visualizing network connectivity and network performance parameters.

Texture Modeling Using Self-Similar Wavelets and POMMs

Author:: Jennifer Davidson (Iowa State University)
Abstract:: Stochastic models are well suited for the modeling of natural texture images. The wavelet transform has also been useful in analyzing texture data. While previous work has supported this, a factor limiting the widespread use of such models is the high number of parameters necessary to describe the model. Recent work has shown that the number of these parameters can be reduced by making assumptions about the data, such as self-similarity of the wavelet coefficients and other assumptions. We make further assumptions about the wavelet transform (WT) coefficients in this work to include stationarity of the WT coefficients as well as dependence within and across scales of the WT coefficients. This is done by fitting an empirical version of a stationary partially ordered Markov model (POMM) in the transform domain both within and across scales. The POMM assume dependence of the WT coefficient values in a local neighborhood in scale, and dependence of the parent node in the adjacent scale with coarser resolution. An empirical pdf for the POMM is found that is based on a quantized version of the WT coefficients. The quantization into K states is performed by modeling the WT coefficients with a mixture of two Gaussians and then estimating the parameters of the Gaussians (the variance only, as we assume a zero mean). We use an expectation-maximization algorithm to find the variances, as the states are considered unobserved data. Maximum likelihood is used to determine the K states. Once the states are found, the empirical POMM pdf is calculated, and this is the model used to represent the data. Synthesis is straightforward and real-time since the POMMs have an algorithm that produces a sample of its distribution with a single visit to each of the pixel locations in the image. Examples are shown of the synthesis technique for varying values of K and neighborhood size for the POMM. Other applications are also being pursued, including classification of texture data and encoding for storage and transmission.

The Adaptive Data Cube: An Experiment in Hyperspectral Pattern Recognition

Author:: Carey E. Priebe (Johns Hopkins University)
Abstract:: We present an experiment in hyperspectral pattern recognition designed to illustrate the potential of the ``adaptive data cube'' in integrated sensing and processing (ISP).

Visualization for Data Mining

Session Organizer: Diane Cook
Session Time: Thursday, June 14th, 2:00pm - 3:45pm
Room Location: Santa Ana Room
Session Chair: Diane Cook, Iowa State

GGobi Meets R: An Extensible Environment for Interactive Dynamic Data Visualization

Author:: Deborah F Swayne (AT\&T Labs - Research)
Abstract:: GGobi is a direct descendant of XGobi, designed so that it can be embedded in other software and controlled using an API (application programming interface). This design has been developed and tested in partnership with R. When GGobi is used with R, the result is a full marriage between GGobi's direct manipulation graphical environment and R's familiar extensible environment for statistical data analysis. GGobi has several other advances over XGobi, including multiple plotting windows, more flexible color management, xml file handling, and portability to Windows.

Graphical Post-Analysis of Association Rules

Author:: Heike Hofmann (University of Ausgburg, Germany)
Abstract:: Association Rules are a widely used tool in data mining. Major problems originate from the mass of output as well as from the restriction to support and confidence for measures of quality. We will introduce graphical techniques for examining association rules. These allow us not only to assess the quality of a single rule visually, but they also provide an overview of the structure among the rules, laying the basis for an interpretation and extraction of ``real'' results.

Uncovering Complexity in Data through Sound

Author:: Mark H. Hansen (Bell Laboratories)
Abstract:: Today, almost every aspect of our lives can be ``rendered'' digitally. Advances in data collection technologies have made commonplace continuous, high-resolution measurements of our physical environment (weather patterns, seismic events, ecological indicators). Equally open to observation are our routine movements through and interactions with our physical surroundings (automobile and air traffic, large-scale land use). In computer-mediated settings, our activities either depend crucially on or consist entirely of complex digital data (financial transactions, accesses to global information systems, web site and internet usage). As a reflection of the diversity and variety of the systems under study, these data-based descriptions of our daily lives tend to be massive in size, dynamic in character, and replete with rich structures. The advent of these enormous repositories of digital information presents us with an interesting challenge: how can we represent and interpret such complex, abstract and socially important data? In a new collaboration, we have begun to give voice to a variety of internet-related data streams. In our initial work, we studied traffic across the web site www.lucent.com. Recently, our focus has been on capturing large-scale ``chatter'' on the web. We have built monitoring agents that allow us to collect streams from thousands of public forums, bulletin boards and chat rooms. The incorporation of textual components in our audio displays presents new and interesting challenges. In this talk, we will try to put our work in context by presenting a brief (and biased) review of the use of data in music compositions, as well as previous attempts to incorporate sound directly into the process of data analysis. This work is part of a Lucent program sponsoring collaborations between researchers at Bell Laboratories and the Brooklyn Academy of Music.

Beyond Correlation

Session Organizer: Thomas Belin
Session Time: Thursday, June 14th, 2:00pm - 3:45pm
Room Location: Costa Mesa Room
Session Chair: Thomas Belin, UCLA

Causal Inference in Statistics: A Gentle Introduction

Author:: Judea Pearl (UCLA)
Abstract:: This talk will provide a conceptual introduction to causal inference, aimed at assisting researchers gain access to recent advances in this area. The talk will stress the paradigmatic shifts that must be undertaken in moving from traditional statistical analysis to causal analysis of multivariate data. Special emphasis is placed on the assumptions that underly all causal inferences, the languages used in formulating those assumptions, and the conditional nature of causal claims inferred from nonexperimental studies. These emphases will be illustrated through a brief survey of recent results, including the control of confounding, corrections for noncompliance, and a symbiosis between counterfactual and graphical methods of analysis. Background information can be viewed on \path|http://www.cs.ucla.edu/~judea/|.

The Defining Role of ``Principal Effects" in Comparing Treatments Using General Post-Treatment Variables

Author:: Constantine E. Frangakis (The Johns Hopkins University)
Abstract:: We study the general and demanding problem of how to make treatment comparisons that adjust for a post-treatment variable, because standard methods can create post-treatment selection bias precisely in situations where there is a scientific reason for the adjustment. We propose a general framework for comparing treatments adjusting for post-treatment variables that yields ``principal effects'' based on ``principal stratification''. Principal stratification with respect to a post-treatment variable is a cross-classification of subjects defined by the joint potential values of that variable under each of the treatments being compared. Principal effects are defined as causal effects within a principal stratum. The key property of principal strata is that they are not affected by treatment assignment and, therefore, can be used as a pre-treatment covariate in defining any covariate-based estimand. As a result, the central property of our principal effects is that they are always causal effects, and do not suffer from post-treatment selection bias. We discuss briefly that principal causal effects are the link between two recently worked applications with post-treatment variables on: (i) treatment noncompliance; and (ii) missingness of outcomes (dropout) following treatment noncompliance. We then discuss the open problem with surrogate endpoints, where we show, using principal effects, that all current definitions of surrogacy, even when perfectly true, cannot generally have interpretation as causal effects attributable to the surrogate. We also formulate a new approach based on principal stratification and principal effects, and show that it has better properties than the standard methods.

Statistical Models for Text

Session Organizer: John Lafferty
Session Time: Thursday, June 14th, 2:00pm - 3:45pm
Room Location: Viejo Room
Session Chair: Andrew McCallum

Active Learning for Support Vector Machines with Applications to Text Classification

Author:: Simon Tong (Stanford University)
Abstract:: Support vector machines have met with significant success in numerous real-world learning tasks. However, like most machine learning algorithms, they are generally applied using a randomly selected training set classified in advance. In many settings, we also have the option of using pool-based active learning. Instead of using a randomly selected training set, the learner has access to a pool of unlabeled instances and can guide the sampling process by querying for the labels of certain pool instances based upon the data that it has seen so far. We introduce a new algorithm for performing active learning with support vector machines, i.e., an algorithm for choosing which instances to request next. We provide a theoretical motivation for the algorithm using the notion of a version space. We present experimental results showing that employing our active learning method can significantly reduce the need for labeled training instances in both the standard inductive and transductive settings.

Conditional Random Fields for Text Processing

Author:: John Lafferty (Carnegie Mellon University)
Abstract:: We present a framework for building probabilistic models to segment and label sequence data using random fields that are globally conditioned on the input sequence. Conditional random fields offer several advantages over hidden Markov models and stochastic grammars for such tasks, including the ability to relax the strong independence assumptions made in those models, and to incorporate hierarchical and overlapping features into the model. Conditional random fields also avoid a fundamental limitation of maximum entropy Markov models (MEMMs) and other discriminative Markov models based on directed graphical models, which can be biased towards states with few successor states. We present iterative parameter estimation algorithms for conditional random fields and compare their performance to HMMs and MEMMs on synthetic and natural language data.

Relevant Encoding of Linguistic Data via the Information Bottleneck Method

Author:: Naftali Tishby (The Hebrew University of Jerusalem)
Abstract:: We introduce a general information theoretic method for extracting relevant information from one set of variables about another, relevant, set. The mutual information between two random variables answers the question: `what is the minimal number of yes/no questions (bits) that are needed to be asked about the variable X in order to learn all we can about the variable Y?' This value does not tell you anything, however, on the content of these questions. What is it that we need to know about the variable X that provides information about the variable Y? We call (the answers to) these questions the relevant components (information) in X about Y, and propose a general method for generating such questions - relevant encoding of X with respect to Y. We introduce both top-down and bottom-up algorithms for this task and discuss their applications for data clustering, text categorization and classification, word-sense disambiguation, and Bioinformatics. Based on joint work with Noam Slonim, Bill Bialek, and Fernando Pereira.

Bayesian Methods (Contributed Session)

Session Time: Thursday, June 14th, 2:00pm - 3:45pm
Room Location: Capistrano Room
Session Chair: S. James Press, UC Riverside

A Split-Merge Markov Chain Monte Carlo Procedure for the Dirichlet Process Mixture Model

Author:: Sonia Jain (University of Toronto)
Abstract:: We propose a split-merge Markov chain algorithm to address the problem of inefficient sampling for conjugate Dirichlet process mixture models. Traditional Markov chain Monte Carlo methods for Bayesian mixture models, such as Gibbs sampling, can become trapped in isolated modes corresponding to an inappropriate clustering of data points. This article describes a Metropolis-Hastings procedure that can escape such local modes by splitting or merging mixture components. Our Metropolis-Hastings algorithm employs a new technique in which an appropriate proposal for splitting or merging components is obtained by using a restricted Gibbs sampling scan. We demonstrate empirically that our method outperforms the Gibbs sampler in situations where two or more components are similar in structure.

Priors for Bayesian Neural Networks

Author:: Mark Robinson (University of British Columbia)
Abstract:: In recent years, Neural Networks (NN) have become a popular data-analytic tool in Statistics, Computer Science and many other fields. NNs can be used as universal approximators, that is, a tool for regressing a dependent variable on a possibly complicated function of the explanatory variables. The NN parameters, unfortunately, are notoriously hard to interpret. Under the Bayesian view, we propose and discuss prior distributions for some of the network parameters which encourage parsimony and reduce overfit, by eliminating redundancy, promoting orthogonality, linearity or additivity. Thus we consider more senses of parsimony than are discussed in the existing literature. We investigate the predictive performance of networks fit under these various priors.

Adaptive Metropolis-Hastings Samplers for the Bayesian Analysis of Large Linear Gaussian Systems

Author:: Stephen KH Yeung (University of Newcastle upon Tyne)
Abstract:: This paper concerns the implementation of efficient Bayesian computation for large linear Gaussian models containing many latent variables. Such models often arise in the context of dynamic linear modeling, where the underlying stochastic process evolves through time, and in more conventional applications such as in hierarchical linear modeling.

Genetic Analysis of Melanoma Onset by Using Estimating Equations and Bayesian Hierarchical Models

Author:: Kim-Anh Do (University of Texas M.D. Anderson Cancer Center)
Abstract:: There are complex relative contributions of genetic and shared environmental factors to an increased risk in melanoma. Data from the Queensland Familial Melanoma Project comprising 15,907 persons from the 1,912 families of 2,118 melanoma cases were analyzed to estimate the additive genetic, common and unique environmental contributions to variation in the age at onset of melanoma. Two complementary approaches for analyzing correlated time-to-onset family data were considered: the generalized estimating equations (GEE) method in which one can estimate relationship-specific dependence simultaneously with regression coefficients that describe the average population response to changing covariates; and a subject-specific Bayesian mixed model in which heterogeneity in regression parameters is explicitly modeled and the different components of variation may be estimated directly. The proportional hazards and Weibull models were utilized, as both produce natural frameworks for estimating relative risks while adjusting for simultaneous effects of other covariates. A simple Markov Chain Monte Carlo method for covariate imputation of missing data was used and the actual implementation of the Bayesian model was based on Gibbs sampling using the free ware package BUGS. In addition, we also used a Bayesian model to investigate the relative contribution of genetic and environmental effects on the expression of naevi and freckles, which are known risk factors for melanoma.

GDAGsim: Sparse Matrix Algorithms for Bayesian Computation

Author:: Darren J. Wilkinson (University of Newcastle)
Abstract:: GDAGsim is a C software library for analysis of conditionally specified linear models. In particular, it can be used to carry out conditional sampling of Gaussian Directed Acyclic Graph (GDAG) models, and hence can be used for the implementation of efficient block MCMC samplers for such models.

Army Research Office Overview

Session Organizer: Robert Launer
Session Time: Thursday, June 14th, 2:00pm - 3:45pm
Room Location: Laguna Room
Session Chair: Robert Launer, ARO

Approximations to Dirichlet Processes with Applications

Author:: Jayaram Sethuraman (Florida State University)
Abstract:: Dirichlet process, which can be used as priors for an unknown distribution appearing in the modeling of data, was introduced by Ferguson. A direct constructive definition of a Dirichlet process given by us simplifies the proofs of many properties. In this talk we show several approximations to Dirichlet processes which will be useful in computational Bayesian analysis involving Dirichlet processes. We present proofs to show that these approximations work in appropriate senses adequate for applications. Finally we present examples of applications of these approximations in some computational Bayesian problems.

Banks of Interacting Bayesian Filters

Author:: Boris L. Rozovskii (University of Southern California)
Abstract:: Emerging applications of statistics to network intrusion detection, estimation of network congestion, target tracking, and volatility estimation in financial markets, etc. bring to the forefront new, more complicated problems in state estimation. These applications require models with abrupt sporadic changes and unusual structures of observation. This paper proposes the use of randomly modulated jump-diffusion systems to model behavior of such systems. We present a bank of interacting Bayesian filters that provide an optimal (in the mean-square sense) estimate for the state process. A recursive algorithm for computing the estimates will be discussed.

Data Reduction by Quantization

Author:: Edward J. Wegman (George Mason University, Center for Computational Statistics )
Abstract:: Massive data sets challenge the limits of both computability and visualization. It is therefore desirable to compress data sets; approximately $10^6$ to $10^7$ Bytes seems desirable. This can be done by sampling (thinning) or quantization (binning). Binning essentially maps the original sample space into a new discrete sample space. It is commonly thought that data is sparse (lumpy) in high dimensions. Binning therefore consists of identifying clusters and determining statistical properties within clusters. Our proposal is to identify statistical properties within bins and replace original data with statistically equivalent data of a much smaller scale. This paper explores these ideas.

The Principle and Practice of Minimum Description Length

Author:: Bin Yu (University of California, Berkeley)
Abstract:: The Minimum Description Length (MDL) Principle for statistical modeling states Occam's razor in the precise language of coding/information theory. It on one hand generalizes the Maximum Likelihood Principle, and on the other motivates useful and effective model selection criteria. Moreover, it serves as an objective platform for comparing model selection procedures from frequentist and Bayesian statistics alike. In this talk, I will give an overview of MDL and describe a fast and low-delay perceptually lossless coder for music/speech based on cascaded LMS and MDL weighting.

Gene Expression (I) (Contributed Session)

Session Time: Thursday, June 14th, 4:15pm - 6:00pm
Room Location: Santa Ana Room
Session Chair: Dennis Kibler, UC Irvine

A Bayesian Approach to the Analysis of cDNA Microarray Data

Author:: M.A. Black (Purdue University)
Abstract:: The recent explosion of interest in microarray technology has resulted in this becoming the preferred methodology for conducting gene expression experiments. Although the ability of an array experiment to examine the expression of thousands of genes simultaneously gives a previously unheard of level of insight to researchers, it also raises a plethora of statistical questions regarding both the sheer volume of data being produced, as well as the level of variability inherent in this relatively new technology. In this paper we present statistical methods based on Bayesian linear models for investigating the various sources of variability present in array experiments. Examples involving data from cDNA microarray experiments conducted at the Purdue University Computational Genomics facility will be used to illustrate this methodology.

A Statistical Analysis of Radiolabeled Gene Expression Data

Author:: Rafael A. Irizarry (Johns Hopkins University)
Abstract:: This paper considers statistical issues in the analysis of a designed experiment to investigate differential gene expression in colon cancer and normal colon tissue. In this experiment gene expression is measured using radiolabeling--based array filters. Specific statistical issues arise in connection with radiolabeling technology, because of the absence of direct control, which are replaced by empty spots on the filter, and with designed experiments, because of the opportunity to systematically quantify important sources of random variation. Here we consider three aspects in detail: normalization of expression intensities; shrinkage estimates of intensity ratios between cancer and normal tissue; and ranking of genes by the strength of the evidence that they are differentially expressed. We propose a robust and simple--to--implement procedures for normalization and shrinkage, that addresses in a technology--specific way the problem of estimating ratios in presence of small and noisy denominators. We also discuss a graphical display to rank genes using a metric based on quantiles of a null distribution obtained by replicating the array experiment in normal tissue.

Replication and Appropriate Statistical Analysis Are Required For Accurate Interpretation of DNA Microarray Experiments

Author:: She-pin Hung (University of California, Irvine)
Abstract:: In its most simple sense, a DNA microarray is defined as an orderly arrangement of hundreds to hundreds of thousands of unique DNA molecules (probes) of known sequence. There are two basic sources for the DNA probes on an array. Either each unique probe is individually synthesized {\it in situ\/} on a rigid surface, or pre-synthesized probes (oligonucleotides or PCR products) are attached to the array platform (usually glass or nylon membranes). Prior to the studies reported here, it was not clear whether comparable data could be obtained by these different DNA microarray formats. Here we report the use of rigorous statistical methods to compare data obtained from {\it in situ\/} synthesized and pre-synthesized DNA microarrays, the Affymetrix GeneChip${}^{\scriptstyle\rm TM}$, and Sigma Genosys${}^{\scriptstyle\rm TM}$ nylon filters, respectively. The results dramatically demonstrate the necessity of replication and appropriate statistical analysis for the interpretation of results of DNA microarray experiments.

Identifying Statistically Significant Similarities in Gene Expression Patterns via Bayesian Infinite Mixture Models

Author:: Mario Medvedovic (University of Cincinnati Medical Center)
Abstract:: The recent development of DNA microarray (DNA ``chip'') technologies for parallel monitoring expression levels of a large number of genes holds a promise of taking our understanding of molecular processes underlying normal functions of living organisms as well as underlying mechanisms of human diseases to a new level. The ability of the DNA microarray technology to produce expression data on a large number of genes in a parallel fashion has resulted in new approaches to identifying individual genes as well as whole pathways involved in performing different biologic functions. One commonly used approach in making conclusions from microarray data is to identify groups of genes with similar expression patterns across different experimental conditions through a cluster analysis. Biologic significance of results of such analyses have been demonstrated in numerous studies. Various clustering procedures, ranging from simple agglomerative hierarchical methods to optimization based global procedures and Self Organizing Maps, have been used for clustering gene expression profiles. In identifying patterns of expression, such procedures depend on either a visual identification of patterns (hierarchical clustering) in a color-coded display or on the correct specification of the number of patterns present in data prior to the analysis (k-means and Self Organizing Maps). We developed a statistical procedure based on the Bayesian Infinite Mixture model in which conclusions about the probability of a set of gene expression profiles being generated by the same pattern are based on the posterior probability distribution of clusterings given the data. In contrast to the Finite Mixture approach, this model does not require specifying the number of mixture components and resulting clustering is a result of averaging over all possible number of mixture components. We implemented a Gibbs sampling based algorithm for generating sample from the posterior distribution of clusterings and used it to identify groups of genes with similar expression profiles in a publicly available dataset.

An Interdisciplinary Program Employing Computational, Biochemical and Genomic Methods to Examine the Effects of Chromosome Structure on the Regulation of Gene Expression

Author:: Lorenzo Tolleri (University of California, Irvine )
Abstract:: We have computed the position of all of the SIDD sites present on the {\it E. coli\/} chromosome at a mid-physiological superhelical density of $\sigma=-0.05$. These calculations are based on a statistical mechanical approach in which the governing partition function and other quantities of interest are evaluated to a high degree of precision. Further computations to determine the positions of all of the SIDD sites on the chromosome at the superhelical densities encountered in topoisomerase mutant strains are currently underway. We are using Hidden Markov Models to predict the location of all the high affinity IHF binding sites on the {\it E. coli\/} chromosome. To determine the {\it in vivio\/} occupancy of IHF at each of its chromosomal binding sites, we are developing a procedure using {\it in vivio\/} IHF-DNA crosslinking, immunoprecipitation and DNA microarrays that we call (CLIP-on-a-CHIP). For these experiments, cells in log phase are treated with formaldehyde to covalently link chromosomal DNA with DNA binding proteins.

Graphical Models

Session Organizer: Michael Jordan
Session Time: Thursday, June 14th, 4:15pm - 6:00pm
Room Location: Costa Mesa Room
Session Chair: Michael Jordan

Introductory Comments

Michael Jordan (UC Berkeley)

Variational Methods and Bayesian Estimation

Author:: Tommi Jaakkola (Massachusetts Institute of Technology)
Abstract:: Sampling methods have typically been used to counter the often prohibitive cost of exact Bayesian calculations. I will discuss an alternative deterministic approach to this problem based on variational methods. Variational methods either generate adjustable simplifying transforms of the likelihood function or operate in the space of distributions and find the best posterior approximation within a simpler family of tractable distributions. The resulting variational transforms are fast and lead to closed form posterior approximations over the parameters and thereby provide an approximate posterior predictive model. The methods can be readily extended to graphical models with complete or incomplete observations with the help of additional variational transforms. I will present the fundamentals of this approach and discuss its limitations and relation to other approaches.

Advanced Mean Field Methods for Probabilistic Models

Author:: Manfred Opper (NCRG, Aston University, Birmingham)
Abstract:: Mean field (MF) methods provide tractable approximations for the computation of high dimensional sums and integrals in probabilistic models. By neglecting certain dependencies between random variables, a closed set of equations for the expected values of these variables is derived which often can be solved in a time that only grows polynomially in the number of variables. This talk deals with principled and general approaches for correcting the deficiencies of simple MF methods which have their origin in Statistical Physics. We also discuss the relation of these methods to belief propagation techniques.

Probability assessment with Maximum Entropy in Graphical Models

Author:

Wim Wiegerinck (SNN - University of Nijmegen)
Tom Heskes (SNN - University of Nijmegen)

Abstract:

The Maximum Entropy (MaxEnt) method is a standard method that searches for the distribution that maximizes entropy under a given set of constraints. Roughly spoken, it selects the distribution that satisfies the given constraints without introducing any additional information. This talk will discuss MaxEnt applied to graphical models. Here, the optimal model has not only to satisfy a given set of constraints, but also a given set of independency to graphical models. Here, the optimal model has not only to satisfy a given set of constraints, but also a given set of independency statements.
As an application, we show how MaxEnt in graphical models can provide a practical tool for the assessment of model parameters in graphical models that are build in collaboration with domain experts.

Environmental Modeling

Session Organizer: Paul Black and Tom Stockton
Session Time: Thursday, June 14th, 4:15pm - 6:00pm
Room Location: Viejo Room
Session Chair: Paul Black

Functional Data Analysis of Complex Computer Simulation Output: A Case Study in nuclear Waste Disposal Risk Assessment

Author:: David Draper (University of California, Santa Cruz)
Abstract:: A key issue in the consolidation process of the nuclear fuel cycle is the safe disposal of radioactive waste. Deep geological disposal based on a multibarrier concept is at present the most actively investigated option (visualize a deep underground facility within which radioactive materials such as spent fuel rods or reprocessed waste, previously encapsulated, are placed, surrounded by other man-made barriers). While the safety of this concept ultimately relies on the safety of the mechanical, chemical and physical barriers offered by the geological formation itself, the physico-chemical behavior of such a disposal system over geological time scales (hundreds or thousands of years) is far from known with certainty. From 1996 to 1999, with partners in Italy, Spain, and Sweden, we were involved in a project for the European Commission, GESAMAC, which aimed in part to capture all relevant sources of uncertainty in predicting what would happen if the disposal barriers were compromised in the future by processes such as geological faulting, human intrusion, and/or climatic change. One major goal of the project was the development of a methodology to predict the radiologic dose for people in the biosphere as a function of time, how far the disposal facility and the other components of the multibarrier system are underground, and other factors likely to be related to dose. For this purpose we developed a complex computer simulation environment called GTMCHEM which ``deterministically'' models the one-dimensional migration of radionuclides through the geosphere up to the biosphere. In this talk I will describe the application of methods of functional data analysis (FDA) to explore the dependence of predicted radiologic dose curves as a function of time on inputs to the computer simulations. FDA includes extensions of traditional statistical methods such as principal components analysis and the analysis of variance (ANOVA) to the case where the outcome, instead of a single real number, is a curve, in our case the logarithm of radiologic dose as a function of the logarithm of time. Previous work in this field was limited to methods such as ANOVA applied to the maximum of such curves; FDA thus permits a much more complete investigation of the relationship between dose and time, and how this relationship depends on the computer simulation inputs.

Integrated Assessment of Drinking Water Regulations

Author:: Mitchell J. Small (Carnegie Mellon University)
Abstract:: The evaluation of the costs and benefits of drinking water regulations in the United States involves multiple disciplines, diverse datasets, and widely differing expectations and beliefs among interested parties. The debate over the EPA's recent proposal (and subsequent withdrawal) of a new maximum contaminant level (MCL) for arsenic highlights the high degree of uncertainty and importance of such an assessment. We present the results of an integrated, Bayesian statistical assessment framework for evaluating the costs and benefits of alternative MCL's for drinking water in the United States. The framework includes: statistical models for the national distribution of contaminant concentrations in (raw) source waters for the $\approx 56,000$ community water suppliers in the US; models for current treatment systems in-place and their pollutant removal efficiencies; and resulting finished-water concentrations, population exposures and health risks. The model also predicts new treatments or other management options that will be adopted in response to new MCL's, their costs, and exposure- and risk-reduction benefits. A full uncertainty analysis is conducted, informed with available data sets for raw-water concentrations, treatments-in-place, finished-water concentrations, and health-risk data. The model is identified and estimated using Bayesian, Markov Chain Monte Carlo methods. The framework is illustrated for a single contaminant (arsenic); and methods are discussed for evaluating multiple MCL's for a suite of contaminants.

Bayesian Sensitivity Analysis and Uncertainty Analysis

Author:: Jeremy E. Oakley (University of Sheffield)
Abstract:: The problem of assessing uncertainties in complex computer simulation codes is of increasing importance in many fields. This is particularly true in environmental applications, where large and highly complex codes are used, where the relevant science is not always clear, and where the implications of errors can be enormous. Although the codes themselves are generally deterministic, statistical methods provide powerful tools for quantifying uncertainties. This talk concerns a range of Bayesian techniques that offer an integrated framework for addressing problems in the use of such models. One major source of uncertainty for the user of a model is often in specifying values for relevant inputs. Running the model may typically mean assigning values to physical parameters which are either unobservable or at least cannot be measured on the scale assumed by the model. Uncertainty in these inputs induces uncertainty on the model outputs, and the objectives of sensitivity and uncertainty analysis are first to quantify this uncertainty and then to explore the role of each uncertain input in the overall output uncertainty. The standard techniques for such analysis involve Monte Carlo methods, which make random draws from the distributions of the inputs, run the code for each sampled input configuration, and thereby obtain a sample of outputs. When the code may take some minutes, hours or even days for a single run, the thousands of runs demanded by Monte Carlo methods become impracticable. The Bayesian tools described in this talk can be applied effectively with far fewer code runs.

Sensitivity Analysis of a Buried Radioactive Waste Risk Model

Author:: Tom Stockton (Neptune and Co)
Abstract:: Complex ecosystem models are useful for investigating dynamics of systems where multiple variables are interacting in a non-linear manner. Quantitatively assessing the importance of input variables becomes more difficult as the dimensionality of the model increases. Sensitivity analysis deals with assigning influence measures to input variables for a given model. Local sensitivity analysis deals with the modification of input parameters one at a time. Although local sensitivity analysis is useful in some applications, the region of possible realizations for the model of interest is left largely unexplored. Global sensitivity analysis attempts to explore the possible realizations of the model more completely. The space of possible realizations for the model can be explored through the use of search curves or evaluation of multi-dimensional integrals using Monte Carlo methods. Sensitivity measures computed by these methods estimate the portion of the total variation of the response that can be attributed to the input variable of interest through an ANOVA-like decomposition of main and interaction effects. Methods of sensitivity analysis such as Multivariate Adaptive Regression Splines (MARS) and the Fourier Amplitude Sensitivity Test (FAST), provide tools that can be used in global sensitivity analysis. To enhance the interpretability of model output by providing a quantitative measure of the importance of input parameters, application of such global sensitivity analysis tools will be demonstrated through an modeling example of involving the performance of buried radioactive waste.

Computational Finance

Session Organizer: Halbert White
Session Time: Thursday, June 14th, 4:15pm - 6:00pm
Room Location: Capistrano Room
Session Chair: Halbert White

Learning to Trade via Direct Reinforcement

Author:: John Moody (Oregon Graduate Institute)
Abstract:: I present new methods for optimizing portfolios, asset allocations and trading systems based on Direct Reinforcement. In this approach, investment decision making is viewed as a stochastic control problem, and strategies are discovered directly. The need to build forecasting models is eliminated, and better trading performance is obtained. The Direct Reinforcement approach differs from dynamic programming and reinforcement algorithms such as TD-learning and Q-learning, which attempt to estimate a value function for the control problem. I present an adaptive algorithm called Recurrent Reinforcement Learning (RRL) that enables a simpler problem representation, avoids Bellman's curse of dimensionality, and offers compelling advantages in efficiency. I demonstrate how Direct Reinforcement can be used to directly optimize risk-adjusted investment returns (including the differential Sharpe ratio), while accounting for the effects of transaction costs. In extensive simulation work, we find that our approach based on RRL produces better trading strategies than systems utilizing Q-Learning (a value function method) or trading based on forecasts. Real world applications include a monthly asset allocation system and an intra-daily currency trader.

Statistical Inference, The Bootstrap, and Neural Network Modeling with Application to Foreign Exchange Rates

Author:: Jeff Racine (University of South Florida)
Abstract:: In this paper we propose tests for individual and joint irrelevance of network inputs. Such tests can be used to determine whether an input or group of inputs ``belong'' in a particular model, thus permitting valid statistical inference based on estimated feedforward neural network models. The approaches employ well known statistical resampling techniques. We conduct a small Monte Carlo Experiment showing that our tests have reasonable level and power behavior, and we apply our methods to examine whether there are predictable regularities in foreign exchange rates. We find that exchange rates do appear to contain information that is exploitable for enhanced point prediction, but the nature of the predictive relations evolves through time.

National Security Agency Overview

Session Organizer: William Szewczyk
Session Time: Thursday, June 14th, 4:15pm - 6:00pm
Room Location: Laguna Room
Session Chair: William Szewczyk, NSA

Dynamic Visualization of Changing Prior and Posterior in Bayesian Analysis

Author:: Hani Doss (Ohio State University)
Abstract:: Over the years, statistical problems given to the NSASAG have often been considered through a Bayesian approach. This invariably raises the question of ``How do you choose the prior?'' Ideally, one would want to know posterior distributions for a wide variety of priors. If the posterior does not change much when one changes the prior then one gets a feeling of reassurance---a different investigator with a slightly different prior may not even bother to recompute the posterior for his prior. On the other hand, if the posterior changes significantly when one changes the prior, then it is important to record that fact, so that for example more time is spent on prior elicitation. Therefore, in almost any problem in which one carries out a serious data analysis, one wants to calculate the posterior distribution for a large number of prior distributions, especially in the exploratory stages of the analysis. In many problems, the posterior is estimated through Markov chain Monte Carlo, which may require non-negligible computer time, and unfortunately this precludes consideration of a large number of priors and an interactive analysis. To deal with this problem, we present a computing environment within which one can interactively change the prior and immediately see the corresponding changes in the posterior. The environment is based on the object-oriented programming language LISP-STAT and an importance sampling procedure which enables one to use the output of one or a small number of Markov chains to obtain estimates of the posterior for a large class of priors. The environment is very general and handles a wide range of standard models, including for example GLM's and hierarchical models.

Nonparametric Clustering

Author:: David W. Scott (Rice University)
Abstract:: The use of density estimation to find clusters in data is supplementing ad hoc hierarchical methodology. Examples include finding high-density regions, finding modes in a kernel density estimator, and the mode tree. Alternatively, a mixture model may be fit and the mixture components associated with individual clusters. Fitting a high-dimensional mixture model with many components is difficult to estimate in practice. Here, we survey mode and level set methods for finding clusters. We describe a new algorithm that estimates a subset of a mixture model. In particular, we demonstrate how to fit one component at a time and how the fits may be organized to reveal the complete clustering model.

Massive Data Sets

Author:

Jon Kettenring (Telcordia Technologies)

Abstract:

TBA

Analyzing Web Data

Session Organizer: Byron Dom
Session Time: Friday, June 15th, 10:45am - 12:30pm
Room Location: Santa Ana Room
Session Chair: Byron Dom

Statistical Learning Problems Associated with the World Wide Web

Author:: Byron Dom (IBM Almaden Research Center)
Abstract:: Problems in statistical machine learning occur in many domains and there is currently much activity in this field. The two major problems addressed are supervised and unsupervised learning, which have also appeared in the context of the world wide web. This talk surveys work on web-specific instances of these two learning problems and others in the hope of making the statistics community more aware of some of the more significant work that has been performed to date. Problems discussed include supervised classification of web pages, clustering of web pages and so-called ``resource discovery'' - finding the most authoritative information on specific subjects.

Finite State Approaches to Information Extraction

Author:: Andrew McCallum (WhizBang! Labs)
Abstract:: Finite state machines are the dominant model for information extraction both in research and industry. In this talk I will give an overview of several finite-state approaches to information extraction, culminating in the presentation of Conditional Random Fields (CRFs), a new model for probabilistic modeling of sequence data. CRFs offer several advantages over hidden Markov models, including the ability to relax strong independence assumptions made in those models. Conditional random fields also avoid a fundamental limitation of maximum entropy Markov models (MEMMs) and other discriminative Markov models based on directed graphical models, which can be biased towards states with few successor states. I will present parameter estimation algorithms several models, as well as experiments with real and synthetic data.

Graph Structure in the Web

Author:: Andrew Tomkins (IBM Almaden Research Center)
Abstract:: This talk discusses recent results concerning the macroscopic structure of the web. We show that the web is a ``bow tie'' decomposable into four equal-sized regions based on connectivity properties. We also examine a number of statistical properties of the web graph, showing results concerning the distribution of in- and out-links, the distribution of sizes of strongly and weakly connected components, and the graph diameter under various definitions. The resulting picture is much less strongly connected than was previously believed.

Bayesian Bioinformatics

Session Organizer: Mike West
Session Time: Friday, June 15th, 10:45am - 12:30pm
Room Location: Costa Mesa Room
Session Chair: Mike West, Duke

Genome-Wide Binding Motif Discovery via Microarray and Prospect Sampler

Author:: Jun Liu (Harvard University)
Abstract:: Recent biological experiments showed that by combining a modified chromatin immunoprecipitation procedure with microarray analysis one can have very strong information on the genome-wide binding locations of a specific tanscription factor (a protein). When combined with a computational strategy such as the Gibbs motif sampler, one can even pinpoint the exact binding motif in some cases. Here we describe an improved algorithm that specifically designed to find the most significant protein-binding motifs when the ChIP-microarray data are available.

Hierarchical Models for Gene Expression Data Analysis

Author:: Michael Newton (University of Wisconsin at Madison)
Abstract:: Hierarchical statistical models provide a flexible approach to the analysis of gene expression data. They enable robust and efficient inference concerning patterns of differential expression among cell types. One rationale for these models is that with thousands of gene-specific questions, it is meaningful to treat the gene-specific parameters as arising from array-specific or cell-type-specific distributions, rather than treating them as fixed effects. Two parametric formulations have proven to be effective: Gamma-Gamma-Multinomial (GGM) and LogNormal-Normal-Multinomial (LNNM) characterize both fluctuations in gene-specific expected expression and fluctuations in measured expression given these underlying means. A nonparametric version of the model provides further insight. I will discuss hierarchical statistical modeling in the context of several experiments involving oligonucleotide chip sets. I will try to demonstrate the utility of reporting the probability of various forms of differential expression, and I will discuss model fitting, model checking, and the role of an interesting arithmetic-geometric mean ratio.

Stochastic Models for Sequences with Non-Local Dependency Structure

Author:: Scott C. Schmidler (Duke University)
Abstract:: We describe a class of probability models for capturing non-local dependencies in sequential data, motivated by applications to biopolymer (protein and nucleic acid) sequence analysis. These models generalize previous work on segment-based stochastic models for sequence analysis. We provide algorithms for Bayesian inference on these models via dynamic programming and Markov chain Monte Carlo simulation. We demonstrate this approach with an application to protein structure prediction.

John Tukey and the Interface

Session Organizer: Karen Kafadar
Session Time: Friday, June 15th, 10:45am - 12:30pm
Room Location: Viejo Room
Session Chair: Karen Kafadar, University of Denver

Does Anyone Know When the Correlation Coefficient is Useful?: A Study of the Times of Extreme River Flows

Author:: David R. Brillinger (University of California, Berkeley)
Abstract:: John Tukey spoke and wrote concerning the uses and limitations of correlation and regression coefficients. In particular he did not think highly of the former, except in limited circumstances and he recognized that there were substantial difficulties of interpretation going along with the latter. The two concepts have random process analogs and this paper considers the case of stationary point processes and their use in a setup that is reasonably well understood physically. The situation is that of the passage of the extreme water flows through a series of locks along the Mississippi River. The focus is on an investigation of the validity of partial coherency analysis, an analog of partial correlation analysis. A maximum likelihood analysis is also presented.

On the Interaction Between Statistics and Computing: In Memory of John W. Tukey

Author:: Luisa Fernholz (Temple University)
Abstract:: This talk is partially based on my unpublished joint work with John W. Tukey. I will discuss some of Tukey's ideas and comments on data analysis, statistics, mathematical statistics, computing, etc. In that context, I will present examples of how the interaction between statistical theory and statistical computing can be used to generate new methodologies that offer powerful tools for analyzing data. In particular, I will present a data dependent method of outlier detection based on the multihalver (the leave-out-half jackknife). The approach is based on the sequences from Plackett-Burman designs with Hadamard matrices used to generate the different ``halvings'' of the data. Examples will be given to show the effectiveness of the multihalver to detect outliers. The multihalver examples are based on my joint work with John W. Tukey, one of his last works in statistics.

The Legacy of John Tukey

Author:: Robert L. Launer (U.S. Army Research Office and the George Washington University)
Abstract:: The talk begins with another look at John Tukey's 1949 paper, ``One Degree of Freedom for Non-Additivity''. This is the spring-board for a look at his career and work. His activity as a consultant to the United States Government is highlighted through personal reminiscence.

Ecological and Earth Science Applications (Contributed Session)

Session Time: Friday, June 15th, 10:45am - 12:30pm
Room Location: Capistrano Room
Session Chair: Christine McLaren, UC Irvine

Spatio-Temporal Prediction of Incomplete Precipitation Records

Author:: Craig Johns (University of Colorado, Denver)
Abstract:: Ecological models depend on preciptation fields as inputs. However, monthly precipitation data from a large number of stations in the United States over a large number of years contain many missing observations. To predict, or infill, we describe a model which does not completely rely on stationary models nor completely upon observed correlations. Modifications to the model are made in order to make the fitting computationally feasible for large data sets.

Bayesian and Frequentist Inference for Ecological Inference: The R x C Case

Author:: Ori Rosen (University of Pittsburgh)
Abstract:: In this paper we propose Bayesian and frequentist approaches to ecological inference, based on $R \times C$ contingency tables, including a covariate. The proposed Bayesian model extends the binomial-beta hierarchical model developed by King, Rosen and Tanner (1999) from the $2 \times 2$ case to the $R \times C$ case. As in the $2 \times 2$ case, the inferential procedure employs Markov Chain Monte Carlo (MCMC) methods. As such, the resulting MCMC analysis is rich but computationally intensive. The frequentist approach, based on first moments rather than on the entire likelihood, provides quick inference via nonlinear least-squares, while retaining good frequentist properties. The two approaches are illustrated with simulated data, as well as with real data on voting patterns in Weimar Germany. In the final section of the paper we provide an overview of a range of alternative inferential approaches which trade-off computational intensity for statistical efficiency.

Using the Chemical Mass Balance Receptor Model to Estimate Pollution Source Contributions from Correlated Air Quality Observations

Author:: William F. Christensen (Southern Methodist University)
Abstract:: In the environmental sciences, receptor models are used to evaluate the contribution of various pollution sources to the air composition at a location. Using pollution source profiles, profile uncertainties, and measurement error variances, the chemical mass balance (CMB) model can be fit in order to partition the ambient pollutants measured at the receptor into a collection of source contributions. We discuss the use of the CMB model for the analysis of a multivariate time series of air quality measurements, and we consider estimation and inference procedures which account for the multiple sources of correlation in the data. Using computer simulation, approaches are compared under various scenarios in which standard model assumptions are violated.

The Application of Ensemble and Combination Classifiers for Land Cover Mapping via Satellite Imagery

Author:: Brian M. Steele (University of Montana)
Abstract:: This talk concerns the application of statistical classification rules for constructing land cover maps from Landsat Thematic Mapper (TM) satellite imagery. These maps are widely used for large-scale management by the USDA Forest Service and other agencies for tasks such as prioritizing fire fighting efforts. The classification problem begins with a Landsat TM scene (approximately $170 \times 170$ km) that has been partitioned as a set of approximately 800,000 polygons. A training set (2000 to 5000 observations) is collected by ground visitation, and a classifier is constructed from the training set to assign land cover type (15-20 types) to the unsampled polygons using the satellite imagery. Good map accuracy is difficult to achieve in mountainous and forested landscapes because land cover type transitions are sometimes indistinct, and because of satellite measurement error. This talk addresses the application of ensemble methods (boosting and bagging), and combination methods for improving classifier accuracy. Two classifiers are of particular interest. I discuss a k-nearest neighbor (k-NN) classifier which estimates group membership probabilities using the exact analytic bootstrap expectations of a k-NN probability estimator. In other words, this classifier amounts to the classifier that would be obtained from averaging all possible bagging versions of a k-NN classifier. The second classifier is a simple spatial classifier that uses the distance from polygons to training observations to classify polygons. While the accuracy of this classifier is quite poor, combinations of spatial and k-NN or tree classifiers are substantially better than any of the constituent classifiers. The method of combining classifiers is not particularly important; in fact, for these data, a very simple method works nearly as well as any.

Mining for Knowledge About Ostracode Assemblages in the Tecolutla River Delta

Author:: A. Dale Magoun (The University of Louisiana at Monroe)
Abstract:: Sediment surface samples containing ostracodes from twenty-one locations representing different aquatic habitats were obtained from a recent study pertaining to the depositional variability of the Tecolutla River delta, Mexico (Chen, Stanley, and Wright, 2000). Aquatic habitats such as an estuary, offshore, a mangrove, tidal ponds, a marsh, and the river were surveyed. The surface samples were analyzed for organic matter and grain size distributions. The ostracodes were identified to their lowest taxonomic level. This paper shows the results of several approaches to the analysis and interpretation of this multivariate study. The methods of this paper provide different approaches to the analysis and the interpretation of the interdependent structures that exists in the aquatic habitats found in this riverine environment. This paper discusses these findings from a traditional taxonomic viewpoint to the latest techniques using correspondence analysis.

International Association for Statistical Computing Overview

Session Organizer: Edward Wegman
Session Time: Friday, June 15th, 10:45am - 12:30pm
Room Location: Laguna Room
Session Chair: Ed Wegman, GMU

Developing Data Mining Systems

Author:: Arno Siebes (Utrecht University)
Abstract:: Data mining is the search for patterns in (large) databases. In the last couple of years I've been involved in the development of two data mining systems and I'm about to start the development of a third system (geared towards bioinformatics). In this talk I will tell you some of my experiences and how they influence the design of the new system. I will especially address efficiency issues related to the interrogation of large databases. For example, what data structures in the database speed-up the query processing and what type of query operators are necessary.

Graphical and Statistical Pruning of Association Rules

Author:: Adalbert Wilhelm (University of Augsburg)
Abstract:: Association rules are amongst the most important patterns that can be discovered using data mining. Their discovery is supported by most, if not all, data mining tools. The analysis and the interpretation of the discovered rules is more difficult or almost impossible, given the huge number of generated rules. In this paper we propose graphical aids and statistical tests to overcome the main drawbacks of mining association rules: arbitrary thresholds for support and confidence, huge number of association rules, meaningless associations due to the presence of frequent itemsets and empirical evaluation of the implications strength. We show how Double Decker plots can be used to visualize association rules. These plots visualize the contingency table that yields the association rule as well as the other potential rules in that table, whether they meet the thresholds or not. This gives a deeper understanding on the nature of the correlation between the left-hand side of the rule and the right-hand side.

*** CANCELLED ***
~~Analyzing High Dimensional Online Monitoring Data~~

Author:: Ursula Gather (University of Dortmund)
Abstract:: In technical process control but also in modern intensive care we are confronted with data structures which are massive as well as high dimensional, dynamic and coming along with complex dependencies between the components at the same time. We report on problems and challenges for statistical data analysis in this situation especially w.r.t. the ability of methods to work online. Some first results concerning dimension reduction and online pattern detection will also be presented.

How Large Is the Web?

Session Organizer: Stephen Fienberg and C. Lee Giles
Session Time: Friday, June 15th, 2:00pm - 3:45pm
Room Location: Santa Ana Room
Session Chair: Andrew Tomkins, IBM Almaden

Searching the Web: Current Limitations, New Techniques, and Future Directions

Author:: C. Lee Giles (Pennsylvania State University)
Abstract:: The World Wide Web continues to revolutionize communication and information systems. Measurements of its size and content provide us with insights into the Web's future and suggestions for new Web tools. Our sampling of the web found that the web though large is not as large as some industrial databases. Our sampling of search engines found that search engines index only a fraction of the web, do not index sites equally, and may not index new pages for months. This talk discusses the impact of these results on future Web search and makes suggestions for future Web tools and research. We illustrate new techniques for information access on the Web, by describing two approaches, metasearch illustrated by Inquirus, a content-based metasearch engine, and by researchindex, a niche search engine, which is the largest free full-text index of computer science literature. This is joint work with Steve Lawrence, Kurt Bollacker and Eric Glover.

How Big is the World Wide Web?

Author:: Adrian Dobra (Carnegie Mellon University)
Abstract:: Considerable efforts have been dedicated to the development of sound procedures for assessing the size of the World Wide Web. The problem is compounded by the fact that sampling directly from the Web is not possible. Several groups of researchers have found sampling schemes which consist of running a number of queries on several major search engines. We present a new approach to analyze datasets collected by query-based sampling, founded on a hierarchical Bayes formulation of the Rasch model for multiple capture-recapture methodology. We illustrate the approach using data gathered by Giles and Lawrence in 1997.

Support Vector Machines

Session Organizer: Tommi Jaakkola
Session Time: Friday, June 15th, 2:00pm - 3:45pm
Room Location: Costa Mesa Room
Session Chair: Tommi Jaakola, MIT

A Tutorial on Support Vector Machines

Author:: Bernhard Sch\"olkopf (Biowulf Technologies)
Abstract:: This talk will present a general tutorial on support vector machines (SVMs).
Talk Slides:: [Postscript]

Author:: Bernhard Sch\"olkopf (Biowulf Technologies)
Abstract:

Chernoff Faces the Interface

Session Organizer: William DuMouchel
Session Time: Friday, June 15th, 2:00pm - 3:45pm
Room Location: Viejo Room
Session Chair: William DuMouchel, AT&T

Graphical Representation as a Discipline

Author:: Herman Chernoff (Harvard University)
Abstract:: The advent of the computer has facilitated the development and use of graphical representation. At the same time it has made such use more important. In response to such needs, many novel techniques have originated. It is time to replace naive dogmas about what constitutes a good representation with a serious study of fundamental principles.

Clustering and the Genetics of Complex Disease

Author:: Richard Olshen (Stanford University)
Abstract:: This work is in collaboration with Alfred Lin, Jing Huang, Neil Risch, David Cox, Koustubh Ranade, Yii-Der Ida Chen, Dee Pei, Chii-Min Hwu, David Curb, Beatriz Rodriguez, Victor Dzau, and many others. Part of the talk will be a description of two ongoing projects in which I am involved, both with many other individuals. Each is a search for genes that predispose to what seems to be a polygenic disease. The first project is SAPPHIRe (Stanford, Asian, and Pacific Program in Hypertension and Insulin Resistance), a network of the NHLBI's Family Blood Pressure Program. The other is supported by the Donald W. Reynolds Foundation and is concerned with cardiovascular diseases, in particular with genes expressed in vessel walls. Most of the talk will be about defining phenotypes in general and so-called ``intermediate phenotypes'' in particular. With SAPPHIRe we have applied k-means clustering on variables, none marginally Gaussian, that quantify levels of plasma lipids and the metabolism of insulin and glucose. Some technologies for choosing ``how many clusters'' firmly choose two for the 597 women and somewhat less certainly two for the 535 men. However, others insist upon three clusters. The differences seem of interest, as statistics and as biology. In each case, one cluster is clearly the ``not insulin resistant'' cluster. Cluster membership is significantly associated with hypertension in women, much less so in men. We devised various permutation tests for making inferences about these data, tests that respect family structures.

Mutlivariate Statistical Process Control and Signature Analysis Using Eigenfactor Detection Methods

Author:: Kuang H. Chen (Massachusetts Institute of Technology)
Abstract:: Many businesses now use univariate statistical process control (USPC) in both their manufacturing and service operations. Automated data collection, low-cost computation, product design to facilitate measurement, and demands for higher quality, lower cost, and increased reliability have accelerated the use of USPC. However, in many situations the widespread use of USPC has caused a backlash as processes are frequently adjusted or shut down when nothing is really wrong because the probability of false positives (Type I error) is calculated based on USPC and takes little or no account of the multiple tests that are being performed or the correlation structure that may exist in the data. Attempts to deal with these issues focus on Bonferroni adjustments, Hotelling's T-squared statistics, and the generalized variance. The problem of high dimensionality is commonly addressed by using some form of dimension reduction such as principal component analysis (PCA). Often these methods indicate that some sort of change has taken place, but provide little information about the real nature of that change. In this paper, we develop a multivariate detection method called eigenfactor analysis which combines information contained in the matrix of eigenvectors and the eigenvalues, and is capable of detecting new events and subtle changes in the covariance structure of the process. Information regarding covariance structure in the process can be crucial for feedback, tuning, and control purposes. Unlike univariate cases where the distributions of the null and alternative hypotheses are aligned on the same axis, the orientations of two multivariate distributions can be much more complicated. Multivariate PCA and T-squared techniques project the test samples on a distribution model based on the training data; hence, such techniques assume alignment in the distributions. Consequently, detection strategies that are capable of differentiating directional drifts become very desirable. The paper concludes with an example from semiconductor manufacturing where the goal is to detect the end point in plasma etch using optical emission spectra. Results using the new procedures are then compared with existing univariate and multivariate process control techniques.

Clusters, Outliers, and Density Models (Contributed Session)

Session Time: Friday, June 15th, 2:00pm - 3:45pm
Room Location: Capistrano Room
Session Chair: Bob Newcomb, UC Irvine

Data Sharpening for Higher-Order Density Estimation

Author:: Michael C. Minnotte (Utah State University)
Abstract:: Data sharpening is a method of applying carefully chosen transformations to data to obtain superior properties with simple methods of analysis. For the case of kernel density estimation, we show that transformations based on pilot estimates of the density and its derivatives can lead to arbitrarily high orders of bias reduction in density estimation with second-order (positive) kernels. Although the transformation is bandwidth-dependent, it requires neither subsidiary smoothing parameters nor back-transformation. Unlike estimates generated using traditional higher-order kernels, ours are constrained to be nonnegative. Numerical studies demonstrate that they also have improved mean square error properties and have fewer arbitrary wiggles.

Robust Detection of Multivariate Outliers in High Dimensions and High Levels of Contamination

Author:: Mark Werner (University of Colorado)
Abstract:: Detection of multivariate outliers using classical statistical measures such as the mean and standard deviation is made difficult by the masking effect, where the influence of multiple outliers on these measures limits their identification. Robust methods have been developed which do not use classical measures, but they tend to be computationally prohibitive and are not feasible for use in very large data sets. Building on ideas of Pe{\~n}a and Prieto (2001) and Rocke and Woodruff (2001), we investigate the success of a variety of methods in detecting outliers with particular interest in a mixture of algorithms - using a combination of classical and robust methods to utilize the advantages of each method and thereby achieve greater success than when utilizing the methods separately. We thus investigate the performance of these algorithms on various types of outliers and outlier clusters. We find that different algorithms perform better for different types of outliers and therefore recommend a combination of methods to achieve the highest possible outlier identification rate when confronted with a data set containing unknown outlier types.

The Complexity of the MCD-Problem

Author:: Paul Fischer (University of Dortmund, Lehrstuhl Informatik 2)
Abstract:: In modern statistics the robust estimation of parameters is a central problem. The Minimum Covariance Determinant (MCD), see, e.g.,~[Rous84], is probably the most important robust estimator of multivariate location and scatter. Its algorithmic complexity, however, was unknown and generally thought to be exponential even if the dimensionality of the data is fixed. A number of heuristics for solving the problem have been developed, e.g., Fast-MCD~[Rous99] and Feasible Solution~[HaOl99]. Here we present a polynomial time algorithm for MCD for fixed dimension of the data. In contrast we show that the MCD problem is $NP$-hard if the dimension varies, hence one cannot expect to find efficient algorithms in this case.

Finding Committee Solutions by Clustering Models in Function Space

Author:: Thomas Ragg (University of Karlsruhe)
Abstract:: Forming a committee is an approach for integrating several opinions or functions instead of favoring a single one. Selecting and weighting the committee members is done in several ways by different algorithms. Possible solutions to this problem is still the topic of current research. Our starting point is the decomposition of the committee error into a bias- and variance-like term. Two requests can be derived from this equation: Models should on the one hand be regularized properly to reduce the average error. On the other hand they should be as independent as possible (in the mathematical sense) to decrease the committee error. The first request of regularization can be handled by a Bayesian learning framework. For the second request I want to suggest a new selection method for committee members based on the pairwise stochastic dependence of their output functions, which maximizes the overall independence. Given these pairwise similarity values the models can be separated in classes by a hierarchical clustering algorithm. From the error decomposition of committees I derive a criterion that allows me to find the optimal number of classes, i.e. the optimal stop criteria for the clustering algorithm. The benefits of the approach are demonstrated for committees of neural networks on a noisy benchmark problem as well as on some problems from the UCI repository.

Detection of Novel Samples in Mass Spectral Data Using Cluster Analysis

Author:: Vladimir Svetnik (Merck \& Co., Inc.)
Abstract:: We present an application of cluster analysis to the detection of novel (unusual, outlying) samples in mass spectral data. These samples, unlike the majority of the others, may represent novel chemical structures that are of most interest to the scientists. Details of our specific application are presented in [1]. A typical N by p data set consists of, N, nearly a thousand mass spectral measurements, with intensities measured at p equal to hundreds of mass-to-charge ratios. Since each sample can be represented as a point in the p-dimensional space, search for unusual samples can be considered as a search for the ``abnormal'' (outlier) points in this space. While it is well known that outlier identification and clustering in such high dimensions is extremely difficult, it will be argued that the nature of the mass spectral data does not lend itself to dimension reduction. There is significant attention in the data mining community to the outlier detection problem in large, multidimensional data sets, see for example [2-4]. Similar to [4], our approach utilizes hierarchical clustering to solve this problem. Advantages of this approach are that, unlike many other methods, it places fewer assumptions on the data, and can be used with various distance (similarity) measures. Based on the clustering algorithm, we developed a complete procedure for outlier identification. We start from the working definition of outliers as those samples that are members of clusters with ``small'' cardinalities, and are far away from clusters with ``large'' cardinalities. A parameter, U, which depends on the expected number of novel samples in the data, is used to define ``small'' clusters. We note one monotonic feature of the hierarchical clustering algorithm: the cardinalities of clusters that a sample belongs to is non-decreasing with respect to the number of clusters. This feature of the algorithm allows us to calculate a measure of outlyingness for each sample, kmin, which is the smallest number of clusters at which the sample belongs to an outlying cluster (a cluster with cardinality smaller than the threshold UN). Samples are then ranked by their kmin values, and the top few are submitted for further investigation by scientists. The use of the lowly populated clusters is consistent with the hypothesis that novel samples should represent only a small fraction of the data. We discuss the use of different similarity measures in our method. In particular, we focus on those measures that are traditionally used in the analysis of mass spectral data. We also present a bootstrap procedure that is used to assess the ``confidence'' one should have about the outlyingness of the samples identified by this method.

Journal of Computational Graphics and Statistics Overview

Session Organizer: Andreas Buja
Session Time: Friday, June 15th, 2:00pm - 3:45pm
Room Location: Laguna Room
Session Chair: Andreas Buja

A Computational Approach for Full Nonparametric Bayesian Inference under Dirichlet Process Mixture Models

Author:: Alan E. Gelfand (University of Connecticut)
Abstract:: Widely used parametric generalized linear models are, unfortunately, a somewhat limited class of specifications. Nonparametric aspects are often introduced to enrich this class, resulting in semiparametric models. Focusing on single or k-sample problems, many classical nonparametric approaches are limited to hypothesis testing. Those that allow estimation are limited to certain functionals of the underlying distributions. Moreover, the associated inference often relies upon asymptotics when nonparametric specifications are often most appealing for smaller sample sizes. Bayesian nonparametric approaches avoid asymptotics but have, to date, been limited in the range of inference. Working with Dirichlet process priors, we extend the effort of Gelfand and Mukhopadhyay (1995). In that paper inference was confined to posterior moments of linear functionals of the population distribution. Here, we provide a computational approach to obtain the entire posterior distribution for more general functionals. We illustrate with three applications: investigation of extreme value distributions associated with a single population, comparison of medians in a k-sample problem, and comparison of survival times from different populations under fairly heavy censoring.

Hierarchical Model-Based Clustering For Large Datasets

Author:: Christian Posse (KangarooNet Inc.)
Abstract:: In recent years, hierarchical model-based clustering has provided promising results in a variety of applications. However, its use with large datasets has been hindered by a time and memory complexity that are at least quadratic in the number of observations. To overcome this difficulty, we propose to start the hierarchical agglomeration from an efficient classification of the data in many classes rather than from the usual set of singleton clusters. This initial partition is derived from a subgraph of the minimum spanning tree associated with the data. To this end, we develop graphical tools that assess the presence of clusters in the data and uncover observations difficult to classify. Using this approach, we analyze two large, real datasets: a multi-band MRI image of the human brain and data on global precipitation climatology. In the last case, we discuss ways of integrating the spatial information in the clustering analysis. We focus on two-stage methods, in which a second stage of processing using established methods is applied to the output from the algorithm presented in this paper, viewed as a first stage.

Software Support for Bayesian Analysis Systems

Session Organizer: Wray Buntine and Bernd Fischer and Johann Schumann
Session Time: Friday, June 15th, 4:15pm - 6:00pm
Room Location: Santa Ana Room
Session Chair: Bernd Fischer, NASA Ames

Computing Environments for Bayesian Statistics

Author:: Robert Gentleman (Harvard School of Public Health)
Abstract:: Bayesian computing should be more than the application of standard Bayesian methods to particular problems. It is certainly important to be able to apply this methodology to practical problems, easily. It is perhaps more important to have a platform that encourages the development of new Bayesian methodology and helps us extend the functionality of existing tools. There are relatively few computing environments available for Bayesian computing. In this talk I will discuss some of the functionality that is needed for Bayesian computing and some of the implementations that are available. In addition changes or enhancements to existing packages that will encourage the use of Bayesian methods will be proposed. Some prototypes will be explored.

Stochastic Parameterized Grammars for Bayesian Model Composition

Author:: Eric Mjolsness (JPL)
Abstract:: Bayesian analysis systems typically have some input language for describing probabilistic models upon which exact or approximate inference is to be performed by one or more algorithmic engines. Here it is proposed that many useful generative probabilistic models can appropriately be expressed in the form of stochastic grammars, which recursively generate sets of words with numerical parameters attached. The rules in these stochastic parameterized grammars (SPG’s) each have the power of a Boltzmann probability distribution, suggesting the use of mean field theory and other methods for model inversion. Model composition arises at the level of multiple rules in a grammar, and also at the level of entire grammars called as subroutines to implement a rule in other grammars. This SPG viewpoint raises new possibilities for interacting with subject domain experts to create statistical models and data analysis algorithms, but raises new challenges for the language or system implementor in the areas of mathematical notation, algorithm composition (e.g. using clocked objective functions), and software synthesis.

The Bayes Net Toolbox for Matlab

Author:: Kevin Murphy (UC Berkeley)
Abstract:: The Bayes Net Toolbox (BNT) for Matlab is a software package for directed graphical models. It supports exact and approximate inference, parameter and structure learning, and static and temporal models. It is widely used in academia for teaching and research. The BNT web site receives an average of 300 hits per week. In this talk, I will describe some of the features that distinguish it from other Bayes net software packages, some of the advantages and disadvantages of using Matlab, plus plans for future work. For more details, see http://http.cs.berkeley.edu/~murphyk/Bayes/bnt.html

Massive Data Sets

Session Organizer: Greg Ridgeway
Session Time: Friday, June 15th, 4:15pm - 6:00pm
Room Location: Costa Mesa Room
Session Chair: Matthew Schonlau, RAND

Data Squashing: Constructing Summary Data Sets

Author:: William DuMouchel (AT\&T Shannon Labs)
Abstract:: One of the chief obstacles to effective data mining is the clumsiness of managing and analyzing data in very large files. The process of model search and model fitting often require many passes over a large dataset, or random access to the elements of a large dataset. Many statistical fitting algorithms assume that the entire dataset being analyzed fits into computer memory, restricting the number of feasible analyses. Here we define ``large dataset'' as one that cannot be analyzed using some particular desired combination of hardware and software because of computer memory constraints. There are two basic approaches to this problem: either switch to a different hardware/software/analysis strategy or else substitute a smaller dataset for the large one. Here we assume that the former strategy is unavailable or undesirable and consider ways of constructing a smaller substitute dataset. This latter approach was named data squashing by DuMouchel, Volinsky, Johnson, Cortes and Pregibon (1999) ``Squashing flat files flatter'' [KDD'99 Proceedings]. Formally, data squashing is a form of lossy compression that attempts to preserve statistical information. Suppose that the original or ``mother'' dataset is a matrix $Y$ having $N$ rows or entities and $n$ columns or variables. The squashed dataset is a matrix $X$ having $M$ rows and $n+1$ columns, where $M \ll N$. The extra column in $X$ is a column of weights, $w_i$, $i = 1, \ldots, M$, where $w_i > 0$ and $\sum_i w_i = N$. It is assumed that $M$ is small enough so that $X$ can be processed by the desired hardware/software, and that the software can make appropriate use of the weight variable. The $n$-dimensional distribution of the rows of $X$ weighted by the $w_i$ is intended to approximate the distribution of the rows of $Y$ well enough that statistical analysis of $X$ is an acceptable substitute for the desired analysis of $Y$. A squashing procedure is evaluated by how much more closely modeling of the squashed pseudo-data approximates results from the full data than do the results from a random sample of the same size. Methods for data squashing will be presented and compared.

Exploratory Analysis of Retail Sales of Billions of Items

Author:: William F. Eddy (Carnegie Mellon University)
Abstract:: We report some preliminary analyses of a data set collected over the past year from a grocery chain containing hundreds of stores. Each record in the data set represents an individual item processed by an individual laser scanner at a particular store at a particular time on a particular day. Each record contains additional information such as store department, price, etc. together with identifying information such as the particular checkout scanner and, for some transactions, customer identification. The total data set contains billions of items which can be aggregated into hundreds of millions of transactions for millions of repeat customers. This talk will describe a number of analyses we have undertaken. Some of these have simply focused on ascertaining the ``quality'' of the data while others have been more narrowly focussed on simple questions like ``which pairs of items are most frequently purchased together'' or ``what is the relationship between basket size and number of baskets.'' The sheer size of the data set has forced us to go beyond simple ``data mining'' methods and become involved in ``meta-mining:'' the post processing of the results of basic analyses.

Mining Large Datasets

Author:: Johannes Gehrke (Cornell University)
Abstract:: This talk has two parts. I will first survey recent work in scalable decision tree construction over massive training databases. In the second part, I will address algorithms for mining high-speed data streams.

Census 2000: Lessons for Census 2010

Session Organizer: Nancy Gordon
Session Time: Friday, June 15th, 4:15pm - 6:00pm
Room Location: Viejo Room
Session Chair: Nancy Gordon, U. S. Census Bureau

Technology and the 2010 Census

Author:: Carol M. Van Horn (U.S. Census Bureau)
Abstract:: The success of the 2000 census was due, in part, to the application of new technologies such as image capture, Internet and laptop computers for some data collection activities. In the coming decade the Census Bureau is planning a program which integrates three components: collecting long form data in a national survey, the American Community Survey (ACS); enhancing the Master Address File and TIGER, a geographic database to bring into compliance with GPS coordinates; and a re-engineered 2010 short form census. By the next census in 2010 advances in technology will provide opportunities for further successes. Research and testing will involve handheld computers equipped with GPS for the creation of an initial address list and for use in nonresponse and other field followup activities, enabling field workers to enter responses directly into a computer file. The short form data collection holds promise for expanding the Internet and other electronic reporting options as modes for data collection. This paper describes the opportunities, the benefits, and the challenges the Census Bureau faces using technology in the 2010 census. The expanded use of technology in 2010 will greatly reduce the Census Bureau's reliance on paper questionnaires.

The U.S. Census Bureau's MAF/TIGER System, Internal and External Interfaces

Author:: Robert Marx (U.S. Census Bureau)
Abstract:: The U.S. Census Bureau's overall mission is to be the preeminent collector and provider of timely, relevant, and quality data about the people and economy of the United States. To accomplish this mission, the Census Bureau has been using the Master Address File and the Topologically Integrated Geographic Encoding and Referencing (MAF/TIGER) system for more than fifteen years to support its various census and sample survey activities. In addition, the MAF/TIGER database has been used as the foundation of the burgeoning geographic information system (GIS) industry in the United States to support the analytical programs and GIS activities managed by other federal agencies, numerous state, local, and tribal governments, the private sector, and academic organizations. The MAF/TIGER system is an aging national resource. The Census Bureau needs to prepare for significantly more automation in the 21st Century, including: improvement to street and other map feature locations; addition of accurate housing unit locations, and enhanced feature change detection methodology to provide more timely updates; and modernizing the processing environment from ``home grown'' systems to one based on COTS and GIS software. Accomplishing the needed improvements will increase the effectiveness of the sponsoring and participating organizations that depend on the Census Bureau's statistical data and its geographic infrastructure.

(Title and abstract unavailable)

Author:: Latanya Sweeney (Carnegie Mellon University)
Abstract:

Gene Expression (II) (Contributed Session)

Session Time: Friday, June 15th, 4:15pm - 6:00pm
Room Location: Capistrano Room
Session Chair: MaryAnn Hill, UC Irvine

Assessing Patient Survival Using Microarray Gene Expression Data Via Partial Least Squares Proportional Hazard Regression

Author:: Danh V. Nguyen (UC Davis)
Abstract:: High dimensional data sets from microarray experiments where the number of variables (genes) $p$ far exceed the number of samples $N$ render most traditional statistical tools of little direct use. However, some of these statistical tools when used in conjunction with an appropriate dimension reduction method can be effective. In this paper we introduce the use the proportional hazard (PH) regression (Cox 1972) in conjunction with dimension reduction by partial least squares (PLS), since the number of covariates $p$ exceeds the number of samples $N$. This setting is typical of gene expression data from DNA microarrays. Specifically, for a given vector of response values which are times to event (death or censored times) and $p$ gene expressions (covariates) we address the issue of how to assess (estimate) the survival experience (curve) when $N \ll p$. The approach taken to cope with the high dimensionality is to reduce the dimension via some dimension reduction (component extraction) method in the first stage and then estimate the survival distribution using a PH regression model in the second stage. The primary method of component extraction considered is PLS. PLS achieves dimension reduction by constructing components to maximize the covariance between the response (survival times) and the linear combination of the covariates (gene expressions) sequentially. This is analogous to principal components analysis (PCA) but the optimization criterion in PCA is variance rather than covariance in PLS. We demonstrate the use of the methodology to a diffuse large B-cell lymphoma (DLBCL) textit{complementary} DNA (cDNA) data set.

Lessons Learned From Analyzing the Differential Gene Expression Data Between Normal and Tumor Tissues in Head and Neck Cancer Patients

Author:: J. Jack Lee (University of Texas M.D. Anderson Cancer Center)
Abstract:: Gene expression in head and neck cancer patients was assessed by using the Research Genetics cDNA membranes GF200 and GF211. The major objective is to identify differentially expressed genes between normal and tumor tissues. In the presentation, we will share many lessons we have learned in acquiring, displaying, and analyzing the microarray data. Before the data analysis, experimental condition need to be carefully documented. For example, information on patient characteristics, tissue extraction method, choice of primer, lot number and strip number of the membrane, hybridization procedure, exposure time, and other parameters of image acquisition all needs to be recorded. We performed duplicate experiments and, in some cases, aligned the images multiple times to estimate the alignment variability, experiment variability, and patient variability. The effect due to multiple stripping of the membrane was also inspected. Standard