Informs Annual Meeting 2017

SA66

INFORMS Houston – 2017

SA65

observe the outcomes corresponding to counterfactual scenarios in observational data, we use techniques from causal inference literature to infer them. We model the problem of learning a decision list as a Markov Decision Process (MDP) and employ a variant of the Upper Confidence Bound for Trees (UCT) strategy which leverages customized checks for pruning the search space effectively. Experimental results on real world observational data capturing judicial bail decisions and treatment recommendations for asthma patients demonstrate the effectiveness of our approach. 2 - Data Science Based Simulation and Optimization of Soybean Variety Selection Durai Sundaramoorthi, Washington University in Saint Louis, St. Louis, MO, 63131, United States, dsundaramoorthi@gmail.com Humanity is facing the greatest challenge of feeding itself. The World Food Programme estimates that about 795 million people do not have adequate food to have a healthy life. About 3.1 million children die every year because of poor nutrition. As a small part of addressing this great humanitarian challenge, this research proposes an analytics framework for growing the optimal soybean vari- eties. Selecting soybean varieties for planting is an important decision that has significant implications for the yield of the farm. With many soybean varieties available along with the data on yield performance collected over several years at various farm locations, the seed-mix selection is a highly intriguing decision to modern day farmers. To aid soybean-mix selection, this research evaluates five data mining models - Regression Trees (RT), Random Forest (RF), Boosted Trees (BT), Multivariate Adaptive Regression Splines (MARS), and Artificial Neural Network (ANN) - as the predictive algorithms for the soybean yield. Using a data set collected between 2008 and 2014 by Syngenta, an agribusiness, we evaluate these models and choose RT as the most suitable algorithm to inform simulation of soybean yield under different weather conditions. We formulated a simula- tion-based optimization problem to determine the optimal soybean-mix to mini- mize the risk associated with the yield. As a result, Syngenta is equipped to make soybean-mix recommendations to the farmers. 3 - Tensor Mixed Effects Model with Applications in Nanomanufacturing Inspection Xiaowei Yue, Georgia Institute of Technology, 755 Ferst Drive NW, ISYE, Atlanta GA, 30332, United States, xwy@gatech.edu” Raman mapping technique has been used to do in-line quality inspections of nanomanufacturing process. In such an application, a massive high dimensional Raman mapping data with mixed effects are generated. In general, fixed effects and random effects in the multi-array Raman data are associated with different quality characteristics such as fabrication consistency, uniformity, defects et al. The existing tensor decomposition methods cannot separate mixed effects, and existing mixed effects model can only handle matrix data instead of high dimen- sional multi-array data. In this paper, we propose a tensor mixed effects (TME) model to handle massive in-line high dimensional Raman mapping data with complex structure. The proposed TME model can (i) separate fixed effects and random effects in a tensor domain; (ii) exploit the correlations along different dimensions; and (iii) realize efficient parameter estimation by a proposed itera- tive double Flip-Flop method. Properties of the TME model and existence and identifiability of parameter estimation are investigated. Numerical analysis demonstrates the efficiency and accuracy of the parameter estimation in the TME model. Convergence and empirical asymptotic properties are discussed in the simulation and surrogate data analysis. The Real case study shows an application of the TME model in quantifying the influence of alignment on carbon nan- otubes buckypaper. 4 - Dynamic Multivariate Functional Data Modeling via Sparse Subspace Learning Chen Zhang, National University of Singapore, Singapore, 117576, Singapore, zhangchen@u.nus.edu Though multivariate functional data are common in many applications, they have not been comprehensively investigated yet, with several challenges to be solved. In particular, they are naturally high-dimensional and have complex cross-correlation structure. Some functions are strongly correlated with similar features, while some others may have almost no correlations with quite diverse features. Furthermore, the cross-correlation structure may even change over time due to the system evolution. With this regard, this paper presents a dynamic subspace learning method for multivariate functional data modeling. In particu- lar, we consider different functions come from different disjoint subspaces, and only functions of the same subspace have correlations with each other. The sub- spaces can be automatically identified and learned by formulating the problem as a sparse regression. By allowing but regularizing the regression change over time, we can describe the cross-correlation dynamics. The model can be efficiently esti- mated by the fast iterative shrinkage-thresholding algorithm (FISTA), and the features of every subspace can be extracted using the smooth multi-channel functional PCA. Numerical studies together with case studies demonstrate the efficiency and applicability of the proposed methodology.

370F Interpretable Machine Learning Sponsored: Data Mining Sponsored Session

Chair: Berk Ustun, Massachusetts Institute of Technology, Cambridge, MA, 02139, United States, ustunb@mit.edu 1 - Optimized Risk Scores

Berk Ustun, Massachusetts Institute of Technology, 20 Highland Avenue 2, Cambridge, MA, 02139, United States, ustunb@mit.edu, Cynthia Rudin Risk scores are simple models to quickly assess risk by adding and subtracting a few small numbers. Such models are widely used in healthcare and criminal justice, but are built ad hoc. We present a new approach to learn risk scores that are optimized for feature selection, integer coefficients, and operational constraints. We formulate the risk score problem as a MINLP, and design a new cutting plane algorithm to solve it efficiently. Our approach can fit optimized risk scores in a way that scales linearly in the sample size, provides a proof of optimality, and addresses complex constraints without parameter tuning. We illustrate its benefits by building customized risk scores for medical applications. 2 - Interpretable and Interactive Approximations of Black Box Models Himabindu Lakkaraju, Stanford University, Escondido Village, 119 Quillen Court, Stanford, CA, 94305, United States, himalv@cs.stanford.edu, Ece Kamar, Rich Caruana, Jure Leskovec We propose Black box Explanations through Transparent Approximations, a novel framework for constructing global explanations of any black-box classifier. We develop a novel objective function which learns a small number of compact decision sets each of which explains black box model behavior in unambiguous, well-defined regions of feature space. Our framework also allows users to interactively explore how the black box model behaves in different subspaces. To our knowledge, this is the first approach which produces explanations of black box classifiers via joint optimization of unambiguity, fidelity, interpretability, while allowing users to explore its behavior based on their preferences. 3 - Interpretable Two-level Boolean Rule Learning Dennis Wei, IBM Thomas J. Watson Research Center, Yorktown Heights, NY, 10598, United States, dwei@us.ibm.com, Guolong Su, Kush R. Varshney, Dmitry M. Malioutov We develop a novel optimization framework for learning two-level Boolean rules for classification, in both conjunctive (AND-of-ORs) and disjunctive (OR-of- ANDs) normal forms. Two objective functions are introduced to trade off classification accuracy and interpretability, where we use 0-1 error and Hamming loss to characterize accuracy and sparsity to characterize interpretability. We propose efficient optimization procedures based on linear programming relaxation, block coordinate descent, and alternating minimization. Experiments show that our algorithms based on Hamming loss provide good tradeoffs between accuracy and sparsity with improvements over the state-of-the-art.

SA66

371A Vehicle Routing with Service Network Design Considerations Sponsored: Transportation Science & Logistics Sponsored Session

Chair: Marco Schutten, University of Twente, Faculty of Management and Governance, P.O. Box 217, Enschede, 7500 AE, Netherlands, m.schutten@utwente.nl Co-Chair: Ali Alyasiry, The University of Queensland, Brisbane, Australia, a.alyasiry@uq.edu.au 1 - A Hub Location and Vehicle Routing Problem with Time Windows Jeremy Castaing, Llamasoft Inc, Ann Arbor, MI, 48104, United States, jeremy.castaing@llamasoft.com, Duyi Li A common extension of the vehicle routing problem (VRP) in real-life transportation system is to model hubs. Hubs are facilities that consolidate shipments between distribution centers and customers. We discuss an approach to extend a classic VRP algorithm to consider hubs and also time windows. The algorithm is integrated in Supply Chain Guru software developed by Llamasoft Inc.

25

Made with FlippingBook flipbook maker