Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SD65

3 - A Deep Learning Approach for Travel Time Prediction Mohammad Abdollahi, Wayne State University, 4815 Fourth Street, Room 1067, Detroit, MI, 48202, United States, Kai Yang Travel time is a fundamental measure in transportation which its accurate prediction is also crucial to the development of intelligent transportation systems. In this paper, we study travel time prediction for New York cabs using deep learning. First, we extract features (by clustering and other techniques) and combine different datasets such as weather and fastest routes. Then, a deep stacked autoencoder for feature representation is presented. These feature transformation makes the feature space more robust and less prone to overfitting. Finally, we compare the effect of feature extraction on performance of different regressors such as boosted trees, deep belief networks, and etc. n SD63 West Bldg 103B Joint Session DM/AI: Data-driven Decision Modeling for Healthcare Sponsored: Data Mining Sponsored Session Chair: Tong Wang, University of Iowa, Iowa City, IA, United States 1 - Feature-efficient Multi-value Rule Sets for Interpretable Patient Mortality Prediction Tong Wang, University of Iowa, Pappajohn Business Build, 21 East Market Street, Iowa City, IA, 52245, United States, Allareddy Veerajalandhar, Sankeerth Rampa, Veerasathpurush Allareddy We propose Multi-vAlue Rule Set (MARS) for predicting patient mortality. Compared to rule sets built from single-valued rules, MARS introduces a more generalized form of association rules that allows multiple values in a condition. Rules of this form are more concise than classical single-valued rules in capturing and describing patterns in data. Our formulation also pursues a higher efficiency of feature utilization, which reduces possible cost in data collection and storage. We applied MARS model on a dataset from Nationwide Inpatient Sample and our model achieved better performance than baseline interpretable models and the patient risk classification system currently used by hospitals. 2 - Optimizing Patient Outcomes via Inverse Classification Michael Lash, University of Iowa, 2 West Washington Street, B4 MacLean Hall, Iowa City, IA, 52242, United States Inverse classification, the process of optimizing the decision features of a test instance using a classifier-based oracle, is a powerful technology that produces personalized, outcome-optimized recommendations. In our formulation, these recommendations are produced by taking into account patient-specific preferences regarding feature priority and cumulative effort. Subsequently, a result is produced showing the improvement in outcome probability and the changes the instance must make to achieve such an improvement. In this talk I present this formulation, along with our corresponding set of methodology, and discuss its application to patient decision making. 3 - Flame – A Fast Large Almost Exact Matching Algorithm for Causal Inference Cynthia Rudin, LSRC / Box 90129, Durham, NC, 27708, United States The FLAME algorithm (Fast Large Almost Matching Exactly) is a large scale matching technique for causal inference. It can handle data so large that it cannot fit in memory, and creates high-quality matches. I will discuss this algorithm and related methods. 4 - Limits of Interpretable Machine Learning in Healthcare Muhammad Aurangzeb Ahmad, University of Washington, Tacoma, WA, United States While interpretability of machine learning systems is critical in holding such systems accountable, practical constraints limit the use of interpretable systems: Comparison of explanations across interpretable machine learning systems, theoretical guarantees of mimic models, soundness vs. completeness of explanations against cognitive limitations, comparison of risk across multiple factors for interpretable models etc. Addressing these limitations will allow us to build better interpretable machine learning systems in healthcare.

n SD64 West Bldg 104A Joint Session DM/Practice Curated: Data Science and Analytics in Healthcare II Sponsored: Data Mining Sponsored Session Chair: Dung Hai Nguyen, Mercy Health, 655 Maryville Center Drive, Saint Louis, MO, 63141, United States 1 - Advances in Density-based Gaze Fixation Identification: Optimization for Outlier Sensitivity, and Automated Detection of Density-modulation Parameter Wen Liu, Worcester Polytechnic Institute, Worcester, MA, United States, Andrew C. Trapp, Soussan Djamabsi Eye tracking is an increasingly common technology with applications to healthcare. Of great interest in eye-tracking studies are fixations, indicative of attention and awareness. However, eye-tracker imprecision can lead to outlier points, e.g. blinks or other anomalies. We extend our density-based fixation identification optimization formulations to account for outlier sensitivity. As our formulations are parameterized by a key density-modulation parameter, we also discuss machine learning approaches for its automatic detection. We conclude with encouraging computational results. 2 - A Continuous Time Bayesian Network Model for Identifying Patterns of Multiple Chronic Conditions United States, Adel Alaeddini, Carlos A. Jaramillo, Mary Jo Pugh Emergence of multiple chronic conditions (MCC) adds complexity in managing patient healthcare design, care, and cost. Hence, it is required to have an effective management of MCC that uses real-time decision making in a big data setting. The proposed study uses de-identified data from a large national cohort of patients (N = 608,503) who entered care in the Department of Veterans Affairs, to identify the risk factors that affect the evolution of MCC. A Continuous Time Bayesian Network is used to examine the interactions of patient disease states and identify major dependencies among MCC that can be used to predict the onset of emerging conditions according to patient level risk factors. n SD65 West Bldg 104B Joint Session DM/Practice Curated: Big Data, Text Mining, and E-commerce Sponsored: Data Mining Sponsored Session Chair: Amarpreet Kohli, University of Southern Maine, P.O. B, Portland, ME, 04104, United States 1 - Dynamic Seed Identification and Activation for Influence Maximization Yerasani Sinjana, Research Scholar, Indian Institute of Technology Kharagpur, Kharagpur, 721302, India, Monalisa Sarma, Manoj Kumar Tiwari In this paper, we consider a network where a set of nodes are termed as seed nodes at each time interval in scheduling seed activation. Seeds are tactically activated for maximizing the spread of influence on social networks: Given a time period for activation, campaign budget, and a network where a set of nodes can be selected as seeds to propagate information. At each stage, time-dependent partial activation of nodes information is used to track the opinions and awareness of users. Activating different users at different periods of time can be termed as Dynamic Seed Activation Problem and can be rewritten as mixed integer programming. A memetic algorithm is employed for scheduling seed activation. Syed Hasib Akhter Faruqui, Graduate Research Assistant, University of Texas-San Antonio, San Antonio, TX, 78256,

113