Informs Annual Meeting 2017

TA65

INFORMS Houston – 2017

TA62

Interactive systems like search engines, recommender systems, and ad placement platforms are ubiquitous. Evaluating and optimizing these systems is hard – users’ feedback governs system performance, and gathering their feedback in randomized experiments is costly. Can we re-use logs collected from deployed systems to reliably perform offline evaluation and learning? I will outline two projects (Viewing Recommendations as Treatments, ICML’16 and De-biasing Learning to Rank, WSDM’17) that advance the state of the art for these problems. 2 - Large-scale Model Validation with Negative Controls Robert Moakler, rmoakler@stern.nyu.edu Causal models are essential for drawing real-world inferences about cause and effect relationships. Therefore, validating these models is crucial, but unfortunately very difficult. Negative control testing, a method typically employed at small-scales, shows promise when applied to large-scale digital applications. In this work, we build a framework for applying negative control testing and argue that the full potential of negative control testing is only realized when companies and organizations are collecting data for many events they observe. We apply this framework to assess and compare RCTs and observational models built for real-world examples in online advertising. 3 - Prophit: Inverse Classification via Causal Deep Learning Michael Lash, University of Iowa, 2 West Washington Street, Past work in the area inverse classification, the process of obtaining instance- specific feature-value updates (recommendations) that optimally improve the probability of a desired classification, focuses on eliciting real-world recommendations, optimization methodology, and classifier inclusion. Ultimately, however, the problem is causal in nature, yet methodologically unexplored. Therefore, we propose Prophit, a deep neural network architecture augmented with causal properties, to solve the inverse classification problem. 4 - Generalized Random Forests Susan Athey, Stanford University, Stanford, CA, United States, athey@susanathey.com, Stefan Wager We propose a method based on random forests (Breiman, 2001) can be used to fit any quantity of interest identified as the solution to a set of local moment equations. Our method operates in covariate space by considering a weighted set of nearby examples. We propose a computationally efficient algorithm for growing generalized random forests, develop a sample theory for our method that shows our estimates are consistent and asymptotically Gaussian and provide an estimator for asymptotic variance. Our approach to develop estimators: non- parametric quantile regression, conditional average partial effect estimation, & heterogeneous treatment effect estimation via instrumental variables. 370F Planning Logistics Systems Under Uncertainty II Sponsored: TSL, Freight Transportation & Logistics Sponsored Session Chair: Mike Hewitt, Loyola University Chicago, Glen Ellyn, IL, 60137, United States, mhewitt3@luc.edu 1 - Covering Path Problem on Grid with an Application to School Bus Routing TA65 Liwei Zeng, Northwestern University, 2145 Sheridan Road, IEMS.Department, Evanston, IL, 60208, United States, liweizeng2015@u.northwestern.edu, Karen Smilowitz, Sunil Chopra We will present work from a collaboration with a large public school district, related to bus transportation to reduce budget pressure. We focus on bus stop selection and route design, two key components of the School Bus Routing Problem identified in the literature. We model the joint problem as a covering path problem. We further exploit the underlying grid-like structure of the road Martijn Mes, University of Twente, Building Ravelijn, P.O. Box 217, Enschede, 7500 AE, Netherlands, m.r.k.mes@utwente.nl, Arturo E. Pérez Rivera We study the problem of selecting transport modes and transfers in a synchromodal network over a multi-period horizon. Freights with different characteristics become known over time. Using probabilistic knowledge of freight arrivals, the planner balances current and future costs at each period. We model this problem as a Markov Decision Process (MDP) and solve it using Approximate Dynamic Programming (ADP) in combination with various exploration strategies. Using different problem settings, we show that our look-ahead approach has significant benefits compared to benchmark heuristics. B4 MacLean Hall, Iowa City, IA, 52242, United States, michael-lash@uiowa.edu, Qihang Lin, Nick Street network and obtain robust and easy-to-implement solutions. 2 - Anticipatory Synchromodal Transportation Planning

370C Statistical Methods for Degradation Data Analysis Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Qingyu Yang, Wayne State University, Detroit, MI, 48202, United States, qyang@wayne.edu Co-Chair: Yong Chen, University of Iowa, Iowa City, IA, 52242-1527, United States, yong-chen@uiowa.edu 1 - Physical-statistical Modeling and Quantification of Degradation Rate for Corroding Aluminum Alloy Xuxue Sun, University of South Florida, 5005 Excellence Blvd 435C, Tampa, FL, 33617, United States, xuxuesun@mail.usf.edu, Mraied Hesham, Wenjun Cai, Guoyuan Liang, Mingyang Li Corrosion-induced degradation of aluminum alloy is a key causing factor for structure failure of many mission critical systems, such as aircrafts. Successful evaluation of degradation performance is important yet challenging since the degradation rate is not directly measurable. We propose a physical-statistical modeling approach in evaluating the alloy degradation rate. Specifically, the latent degradation rate is first characterized via fractional order system dynamics. Bayesian inference is then carried out to address both data uncertainty and population variability due to the limited data samples. A real case study is further provided to illustrate of the proposed work. 2 - Degradation Analysis of High Strength Dual-phase Steel by using Microstructure Images Wujun Si, Wayne State University, 4500 Cass Avenue, Apt 1001, Detroit, MI, 48201, United States, fk9456@wayne.edu, Qingyu Yang, Xin Wu The microstructure of high strength dual-phase steel strongly impacts steel degradation and failure. In this paper, we conduct degradation based reliability analysis of the dual-phase steel by utilizing its microstructure images. A novel functional covariate degradation model is proposed for the analysis. A penalized least square estimation method is developed for the model parameter estimation. Simulation and real case studies are conducted to demonstrate the proposed model. Results show that the failure prediction is more precise by utilizing the microstructure images. 3 - An Integrated Method for Pipeline Fatigue Crack Growth Prediction Mingjiang Xie, University of Alberta, Edmonton, AB, Canada, mingjian@ualberta.ca, Zhigang Will Tian Managing fatigue cracks has been a huge challenge in pipeline industry. An integrated method is proposed for predicting pipeline fatigue crack growth considering large defect sizing uncertainty, which can be used for remaining useful life prediction. Stress analysis models, crack propagation models and in-line inspection (ILI) data are fused in this approach. Model parameters are updated based on ILI data through Bayesian approach to achieve a more accurate crack growth prediction. The proposed method is compared with existing physics-based methods. 4 - A Prognostic Framework for Complex Engineering Systems with Multiple Failure Modes Liexiao Ding, Georgia Institute of Technology, Atlanta, GA, United States, nagi.gebraeel@isye.gatech.edu, Nagi Gebraeel, Xiaolei feng Complex engineering equipment typically undergoes multifaceted degradation processes. The use of multiple sensors potentially allows the capture of different aspects of such complicated degradation processes, which sometimes exhibit different failure modes. In this talk, we present a statistical prognostic framework for multi-failure modes applications. Specifically, a penalized functional data clustering methodology is first built to select the informative sensor signals corresponding to each failure modes. Next, a prognostic model is built by using the informative sensor signals selected earlier. Numerical studies are used to evaluate the performance of our model. 370E Causal Data Mining Sponsored: Data Mining Sponsored Session Chair: Michael Lash, University of Iowa, Iowa City, IA, 52242, United States, michael-lash@uiowa.edu 1 - Building Recommenders and Search Engines using Causal Data Mining Adith Swaminathan, Microsoft, One Microsoft Way, Redmond, WA, 98052, United States, adswamin@microsoft.com TA64

293