INFORMS 2021 Program Book

INFORMS Anaheim 2021

MC01

Monday, 11:00AM-12:30PM

2 - Fair and Interpretable Decision Rules for Binary Classification Connor Lawless, Cornell University, Ithaca, NY, United States, Oktay Gunluk In this talk we consider the problem of building Boolean rule sets in disjunctive normal form (DNF), an interpretable model for binary classification, subject to fairness constraints. We formulate the problem as an integer program that maximizes classification accuracy with explicit constraints on two different measures of classification parity: equality of opportunity, and equalized odds. Column generation framework, with a novel formulation, is used to efficiently search over exponentially many possible rules. When combined with faster heuristics, our method can deal with large data-sets. Compared to other fair and interpretable classifiers, our method is able to find rule sets that meet stricter notions of fairness with a modest trade-off in accuracy. 3 - Learning Fair Optimal Classification Trees Sina Aghaei, University of Southern California, Los Angeles, CA, 90007, United States, Jack Benson, Andres Gomez, Phebe Vayanos The increased use of machine learning (ML) in high stakes domains has created an urgent need for ML algorithms that are fair and interpretable and that leverage the available data to its full extent to yield the most accurate predictions. In this paper, we propose a versatile framework for learning optimal and fair classification trees based on mixed integer optimization technology. Our framework is flexible to capture arbitrary fairness notions from the literature such as statistical parity, conditional statistical parity, etc. We evaluate our method on numerous datasets from the literature and investigate the trade-off between accuracy and fairness. We provide an R package that is freely distributed for academic and non-profit use. 4 - A Statistical Test for Probabilistic Fairness Bahar Taskesen, EPFL, Lausanne, Switzerland Algorithms are now routinely used to make consequential decisions that affect human lives. While algorithms empower us to harness all information hidden in vast amounts of data, they may inadvertently amplify existing biases in the datasets. This concern has sparked increasing interest in fair machine learning. Machine learning models should undergo intensive tests to detect algorithmic biases before being deployed at scale. We use ideas from the optimal transport theory to propose a statistical hypothesis test for detecting unfair classifiers. The test statistic quantifies the distance of the empirical distribution supported on the test samples to the manifold of distributions that render a pre-trained classifier fair. We develop a rigorous hypothesis testing mechanism for assessing the probabilistic fairness of any pre-trained logistic classifier. 5 - Unbiased Subdata Selection for Fair Classification: A Unified Framework and Scalable Algorithms Qing Ye, Virginia Tech, Blacksburg, VA, United States, Weijun Xie Fair classification concerns the biases in the classical machine learning models. Due to high nonconvexity of fairness measures, existing methods often approximate fairness measures via convex programs. This paper fills the gap by developing a unified framework to incorporate fairness measures precisely. In the proposed framework, when the classification outcomes are known, the resulting problem, termed unbiased subdata selection, can be used to enhance the classification fairness by selecting more representative data points. This motivates us to develop an iterative refining strategy (IRS) to improve the classification accuracy and conduct the unbiased subdata selection in an alternating fashion. We prove approximation guarantee of IRS and numerically demonstrate that the proposed framework can yield better fair classification outcomes than existing ones. MC03 CC Ballroom C / Virtual Theater 3 Hybrid APS Special Session on ‘Causal Inference’ Sponsored: Applied Probability Society Sponsored Session Chair: Pengyi Shi, Purdue University, West Lafayette, 47907, United States 1 - Tutorial on Causal Inference in Medicine and Public Health Miguel Hernan, Harvard University, Cambridge, MA, United States This session will provide an introduction to causal inference, including the definition of causal effects, key conditions for their identifiability and methods for their estimation. These concepts will be illustrated with several real world applications to study interventions for the treatment and prevention of disease.

MC01 CC Ballroom A / Virtual Theater 1 INFORMS TutORial Good and Bad Optimization Models: Insights from Rockafellians Tutorial Session Chair: John Gunnar Carlsson, University of Southern California, Los Angeles, 90089, United States 1 - Good and Bad Optimization Models: Insights from Rockafellians Johannes Royset, Naval Postgraduate School, Monterey, CA, 93943-5285, United States A basic requirement for a mathematical model is often that its solution (output) shouldn’t change much if the model’s parameters (input) are perturbed. This is important because the exact values of parameters may not be known and one would like to avoid being misled by an output obtained using incorrect values. Thus, it is rarely enough to address an application by formulating a model, solving the resulting optimization problem and presenting the solution as the answer. One would need to confirm that the model is suitable, i.e., “good,” and this can, at least in part, be achieved by considering a family of optimization problems constructed by perturbing parameters as quantified by a Rockafellian function. The resulting sensitivity analysis uncovers troubling situations with unstable solutions, which we referred to as “bad” models, and indicates better model formulations. Embedding an actual problem of interest within a family of problems via Rockafellians is also a primary path to optimality conditions as well as computationally attractive, alternative problems, which under ideal circumstances, and when properly tuned, may even furnish the minimum value of the actual problem. The tuning of these alternative problems turns out to be intimately tied to finding multipliers in optimality conditions and thus emerges as a main component of several optimization algorithms. In fact, the tuning amounts to solving certain dual optimization problems. In this tutorial, we’ll discuss the opportunities and insights afforded by Rockafellians. MC02 CC Ballroom B / Virtual Theater 2 Hybrid Fair Optimization and Learning under Uncertainty Sponsored: OPT/Optimization Under Uncertainty Sponsored Session Chair: Qing Ye, Virginia Tech, Blacksburg, VA, 24061, United States Co-Chair: Weijun Xie, Virginia Tech, Blacksburg, VA, 24061, United States 1 - A Stochastic Alternating Balance K-means Algorithm for Fair Clustering Suyun Liu, Lehigh University, Smags Apts Lehigh Univ 13 Duh Dr Apt 212, Building, Bethlehem, PA, 18015-3749, United States, Luis Nunes Vicente In the application of data clustering, the clustering outcome might discriminate against people in different demographic groups, leading to unfairness. A natural conflict occurs between the clustering cost and the balance, leading to a nonconvex and nonsmooth biobjective problem. To determine the complete trade-off between the two competing goals, we design a novel stochastic alternating fair k-means (SAfairKM) algorithm consisting of alternating k-means updates and swap updates. Moreover, we propose a novel companion algorithm, the stochastic alternating biobjective gradient descent algorithm, which can handle a smooth version of the biobjective fair k-means problem. A sublinear convergence rate is established under strong convexity for the determination of a stationary point of a weighted-sum function parameterized by the number of updates on each function.

58