Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TD67

2 - A Comparison Between Different Machine Learning Algorithms for Better Accuracy in Trauma Outcomes Prediction Fatima Almaghrabi, PhD Student, The University of Manchester, Booth street, Manchester, United Kingdom, Dong-Ling Xu, Jian- Bo Yang Outcome prediction models are useful in identifying the extent of patient injuries and prioritising immediate life threats. This research aims to identify the most accurate tools for building a prediction model and to increase model accuracy to enhance the care services provided to trauma patients. Thus, the research attempts to identify which algorithms have the highest classification accuracy in predicting trauma outcome. The results of some machine learning (ML) algorithms, such as decision tree, logistic regression, random forest and neural network results were compared to the evidential reasoning rule. 3 - Flexible Job Shop Scheduling with Multi Agent Advantage Actor Critic Jinkyoo Park, Korea Advanced Institute of Science and Technology, 291 Daehak-ro, Yuseong-gu, Daejeon, 34141, Korea, Republic of, Jaehyeong Chun, Jongwoo Ko, Jun Young Park In this research, we propose a multi-agent reinforcement learning approach to schedule a semiconductor manufacturing process. A semiconductor manufacturing process is composed of a large number of sequential jobs conducted by different types of machines. The optimal scheduling policy seeks to minimize the makespan while satisfying various constraints. We formulate the scheduling problem as a multi-agent team game, and then derive the independent policy for each machine by using the multi-agent advantage actor-critic. Experiment results show that our decentralized multi-agent RL approach is effective compared to centralized scheduling approach or heuristic scheduling approach. 4 - Data Driven Sparse System Identification Salar Fattahi, PhD, UC Berkeley, Berkeley, CA, 94702, United States, Somayeh Sojoudi In this work, we study the system identification porblem for sparse linear time- invariant systems. We propose a sparsity promoting estimator to identify the dynamics of the system with only a limited number of input-state data samples. Using contemporary results on high-dimensional statistics, we prove that logarithmic number of data samples is enough to reliably estimate the system dynamics. The developed estimator offers a small estimation error entry-wise and is capable of “exact recovery” of the underlying sparsity structure of the system with small number of data samples. We demonstrate the effectiveness of our approach through different case studies. 5 - Deep Learning in Finance - Estimation of Factor Models Muye Wang, Columbia Business School, 3022 Broadway, Uris Hall, 4H, New York, NY, 10027, United States Factor analysis is a statistical method used to describe variability among observed, correlated variables in terms of a potentially lower number of unobserved variables called factors. By using a variational autoencoder framework, we are able to incorporate outside relevant information to improve linear factor model’s predictive performance. In addition, we also consider extending the traditional linear factor model to a non-linear framework. Finally, we conduct numerical experiments using SP500 daily return data and trading volume data to illustrate the superior performance. Joint Session DM/Practice Curated: Intriguing Tweaks in Data Science I Sponsored: Data Mining Sponsored Session Chair: Özge Sürer, Northwestern University, 1310 Chicago Avenue 3H, Evanston, IL, 60201, United States 1 - A New Fuzzy Set and Quadratic Surface SVM Approach for Mislabeled Classification with Applications Jian Luo, Associate Professor, Dongbei University of Finance and Economics, No. 217 Jianshan Street, Shahekou District, Dalian, 116025, China, Ye Tian, Zhibin Deng To handle mislabeled classification problem, this paper proposes a totally new method which first adopts the intuitionistic fuzzy set method to detect those mislabeled points, then deletes their labels, and utilizes their full position information to build a semi-supervised kernel-free quadratic surface SVM model. Besides, a branch-and-bound algorithm is designed to improve the efficiency and accuracy. Some numerical tests on artificial and real datasets verify the superior performance of proposed method among several benchmark methods. Furthermore, the proposed method is applied to brain-computer interface and credit risk assessment, which strongly demonstrates its effectiveness. n TD66 West Bldg 105A

2 - As Good as it Gets? Upper Bounds on Prediction Performance David Anderson, Villanova University, Philadelphia, PA, United States, Margret V. Bjarnadottir Taking only an assumption that “very similar” observations should have similar predictions to each other, we formulate a linear program to generate upper bounds on the possible predictive accuracy of a given outcome, using data contained in a dataset. We show results on multiple real-world and simulated datasets. 3 - Building Interpretable and Highly Accurate Supervised Learning Models Abdelaziz Berrado, Mohammed V. University, BP 765, Avenue Ibn Sina, Agdal, Rabat, 10080, Morocco Rule based classification algorithms such as CART are very attractive in several applications. They owe their popularity to the simplicity of the tree-building algorithm and the actionability of the resulting models. These algorithms suffer, however, from instability due to the tree building process: they perform greedy searches for rules, which could lead to missing important rules. Ensemble methods can mitigate this weakness of individual tree learners. They result, however, in blackbox models and are unable to provide insights into the structure of the predictive model. We discuss in this work ongoing research aimed at building accurate and highly actionable classification models. 4 - Data Analytics as a Tool for Problem Structuring Patrick Hester, UNC Asheville, 28 Gibson Rd., Asheville, NC, 28804, United States Data analytics is everywhere, from healthcare to business to higher education. Analysts use it to understand, predict, and improve an organization’s performance, as a fundamental tool in one’s Hard OR toolbox. This perspective misses out on an equally valid, yet overlooked, application of analytics to help us frame our problem; in this arena, data analytics can be very powerful. I will argue for the use of data analytics in the problem structuring (or Soft OR) phase, as a natural complement to its continued use during the solution phase of a problem engagement. 5 - Coefficient Tree Regression for Discovering Hidden Structure zge S rer, Northwestern University, 1310 Chicago Avenue 3H, Evanston, IL, 60201, United States, Daniel Apley, Edward C. Malthouse The proliferation of technologies allows us to collect datasets of immense size with a large number of variables. In practice, many groups of predictors often share a common regression coefficient, but the groups are unknown. We propose an algorithm called coefficient tree regression to discover the unknown group structure by utilizing the properties of linear regression in an efficient way. We avoid matrix operations and speed up the computation to obtain an efficient algorithm. Our method achieves high accuracy competitive with existing methods. Finally, we test our algorithm with real datasets and demonstrate that it yields interpretable models by exploring the relations between predictors. n TD67 West Bldg 105B Joint Session ISS/Practice Curated: Mobile App, Commerce, and Analytics Sponsored: Information Systems Sponsored Session Chair: Zhan Shi, Arizona State University, Tempe, AZ, 85251, United States 1 - Are Gaming Apps and Social Media Apps Adversaries or Partners to Education Apps? An Empirical Analysis Sangpil Han, Arizona State University, Tempe, AZ, United States,, Sungho Park, Sanghak Lee, Wonseok Oh This paper examines an empirical question of whether spending time on hedonic apps such as Facebook and Angry Birds promote or impede users’ usage of educational apps. We investigate this question using an individual-level app usage time data set. 2 - Personalized Targeting to Tackle the Challenge of Low Engagement in Mobile Apps: Combining Structural Hidden Markov Model and Field Experiment Yingjie Zhang, Carnegie Mellon University, Pittsburgh, PA, United States Low engagement rates and high attrition rates have been formidable challenges for mobile apps and their long-term success, especially for those whose revenues derive mainly from in-app purchases. This paper proposes a new structural forward-looking Hidden Markov Model (FHMM) and combines it with a randomized field experiment on app notification promotions. Overall, the novel feature of our paper is to propose a new approach of personalized targeting to tackle the challenge of low engagement in mobile apps by combining structural model and field experiment.

371