Research papers
2020 · Princeton University Press eBooks · 11,885 citations
Both in science and in practical affairs we reason by combining facts only inconclusively supported by evidence. Building on an abstract understanding of this process of combination, this book constructs a new theory of epistemic probability. The theory draws on the work of A. P. Dempster but diverges from Depster's viewpoint by identifying his as epistemic probabilities and taking his rule for combining upper and lower as fundamental. The book opens with a critique of the well-known Bayesian theory of epistemic probability. It then proceeds to develop an alternative to the additive set functions and the rule of conditioning of the Bayesian theory: set functions that need only be what Choquet called monotone of order of infinity. and Dempster's rule for combining such set functions. This rule, together with the idea of weights of evidence, leads to both an extensive new theory and a better understanding of the Bayesian theory. The book concludes with a brief treatment of statistical inference and a discussion of the limitations of epistemic probability. Appendices contain mathematical proofs, which are relatively elementary and seldom depend on mathematics more advanced that the binomial theorem.
2017 · arXiv (Cornell University) · 7,624 citations
Understanding why a model makes a certain prediction can be as crucial as the prediction's accuracy in many applications. However, the highest accuracy for large modern datasets is often achieved by complex models that even experts struggle to interpret, such as ensemble or deep learning models, creating a tension between accuracy and interpretability. In response, various methods have recently been proposed to help users interpret the predictions of complex models, but it is often unclear how these methods are related and when one method is preferable over another. To address this problem, we present a unified framework for interpreting predictions, SHAP (SHapley Additive exPlanations). SHAP assigns each feature an importance value for a particular prediction. Its novel components include: (1) the identification of a new class of additive feature importance measures, and (2) theoretical results showing there is a unique solution in this class with a set of desirable properties. The new class unifies six existing methods, notable because several recent methods in the class lack the proposed desirable properties. Based on insights from this unification, we present new methods that show improved computational performance and/or better consistency with human intuition than previous approaches.
2009 · 6,450 citations
Most tasks require a person or an automated system to reason -- to reach conclusions based on available information. The framework of probabilistic graphical models, presented in this book, provides a general approach for this task. The approach is model-based, allowing interpretable models to be constructed and then manipulated by reasoning algorithms. These models can also be learned automatically from data, allowing the approach to be used in cases where manually constructing a model is difficult or even impossible. Because uncertainty is an inescapable aspect of most real-world applications, the book focuses on probabilistic models, which make the uncertainty explicit and provide models that are more faithful to reality. Probabilistic Graphical Models discusses a variety of models, spanning Bayesian networks, undirected Markov networks, discrete and continuous models, and extensions to deal with dynamical systems and relational data. For each class of models, the text describes the three fundamental cornerstones: representation, inference, and learning, presenting both basic concepts and advanced techniques. Finally, the book considers the use of the proposed framework for causal reasoning and decision making under uncertainty. The main text in each chapter provides the detailed technical development of the key ideas. Most chapters also include boxes with additional material: skill boxes, which describe techniques; case study boxes, which discuss empirical cases related to the approach described in the text, including applications in computer vision, robotics, natural language understanding, and computational biology; and concept boxes, which present significant concepts drawn from the material in the chapter. Instructors (and readers) can group chapters in various combinations, from core topics to more technically advanced material, to suit their particular needs.
1989 · Econometrica · 6,031 citations
In this paper, we develop a classical approach to model selection. Using the Kullback-Leibler Information Criterion to measure the closeness of a model to the truth, we propose simple likelihood-ratio based statistics for testing the null hypothesis that the competing models are equally close to the true data generating process against the alternative hypothesis that one model is closer. The tests are directional and are derived successively for the cases where the competing models are non-nested, overlapping, or nested and whether both, one, or neither is misspecified. As a prerequisite, we fully characterize the asymptotic distribution of the likelihood ratio statistic under the most general conditions. We show that it is a weighted sum of chi-square distribution or a normal distribution depending on whether the distributions in the competing models closest to the truth are observationally identical. We also propose a test of this latter condition.
2011 · Statistics in Medicine · 26 citations
This paper describes the problem of public health monitoring for waterborne disease outbreaks using disparate evidence from health surveillance data streams and environmental sensors. We present a combined monitoring approach along with examples from a recent project at the Johns Hopkins University Applied Physics Laboratory in collaboration with the U.S. Environmental Protection Agency. The project objective was to build a module for the Electronic Surveillance System for the Early Notification of Community-based Epidemics (ESSENCE) to include water quality data with health indicator data for the early detection of waterborne disease outbreaks. The basic question in the fused surveillance application is 'What is the likelihood of the public health threat of interest given recent information from available sources of evidence?' For a scientific perspective, we formulate this question in terms of the estimation of positive predictive value customary in classical epidemiology, and we present a solution framework using Bayesian Networks (BN). An overview of the BN approach presents advantages, disadvantages, and required adaptations needed for a fused surveillance capability that is scalable and robust relative to the practical data environment. In the BN project, we built a top-level health/water-quality fusion BN informed by separate waterborne-disease-related networks for the detection of water contamination and human health effects. Elements of the art of developing networks appropriate to this environment are discussed with examples. Results of applying these networks to a simulated contamination scenario are presented.