Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SA67

3 - Diversity Based Link Recommendation Algorithm Kexin Yin, University of Delaware, 1020 Wharton Drive, Newark, DE, 19711, United States By 2018, more than 4 billion people are reported using the internet all over the world, while more than 3 billion people are actively using social media. Social network sites are playing important roles in this information era. Link recommendation, the core technique targeting to help users establish a good friendship network, attracts significant attentions from both industry and academic researchers. However, friends recommended by an existing method are generally homogeneous in terms of their backgrounds, e.g. location or current university, because existing methods recommend friends with similar backgrounds or interests or sharing many common friends with a user. Serving as an important information source for users, social networks can perform better if they can introduce diversity in link recommendation. Improving social network diversity could increase the possibility for users to be connected with different information communities, thus gaining users social capital and competence. Because different users may have different preference of diversity when they make friends, and people can have different diversity preference on different dimensions of user background, link recommendation can be made to have different diversity level according to user’s diversity preference on each user background dimension. In this presentation, multi-dimensional diversity preference is introduced as a new factor in link recommendation, and a new link recommendation problem, namely diversity preference-aware link recommendation problem is defined and proofed to be a NP-hard problem. An efficient heuristic solution method will be developed based on the heuristic algorithm of binary quadratic programing and empirically tested with a real social network data set from Google+.. 4 - Using Long Short-Term Memory to Predict Hospital Readmission Jiaheng Xie, University of Arizona, 1130 E. Helen Street, Tucson, AZ, 85721, United States Hospital readmission refers to the situation where a patient is re-hospitalized with the same primary diagnosis within a specific time interval after discharge. Hospital readmission causes $26 billion preventable expenses to the U.S. health systems annually and often indicates suboptimal patient care. To alleviate those severe financial and health consequences, it is crucial to proactively predict patients’ readmission risk. Such prediction is challenging because the evolution of medical events (illness trajectory) is dynamic and complex. The state-of-the-art studies apply statistical models which assume homogeneity among all patients and use static predictors in a period, failing to consider patients’ heterogeneous illness trajectories. Our approach û TADEL (Trajectory-BAsed DEep Learning) û is motivated to tackle the problems with the existing approaches by capturing various illness trajectories and accounting for patient heterogeneity. We evaluated TADEL on a five-year national Medicare claims dataset including 3.6 million patients per year over all hospitals in the United States, reaching an F1 score of 0.867 and an AUC of 0.884. Our approach significantly outperforms all the state- of-the-art methods. Our findings suggest that health status factors and insurance coverage are important predictors for readmission. This study contributes to IS literature and analytical methodology by formulating the trajectory-based readmission prediction problem and developing a novel deep-learning-based readmission risk prediction framework. From a health IT perspective, this research delivers implementable methods to assess patients’ readmission risk and take early interventions to avoid potential negative consequences. n SA67 West Bldg 105B Joint Session ISS/Smart Cities: Economics of Smart Cities Sponsored: Information Systems Sponsored Session Chair: Beibei Li, Carnegie Mellon University, Pittsburgh, PA, 15213, United States 1 - Learning Individual Behavior Using Sensor Data: The Case of GPS Traces and Taxi Drivers Beibei Li, Carnegie Mellon University, 5000 Forbes Ave, Hamburg Hall 3026, Pittsburgh, PA, 15213, United States Abstract not available. 2 - Peer-to-peer Transportation Platforms and Local Consumer Mobility Zhe Zhang, Carnegie Mellon University, 4800 Forbes Avenue, Pittsburgh, PA, 15213, United States Abstract not available. 3 - Pricing Efficiently in Designed Markets: Evidence from Ride-Sharing John Horton, New York University, New York, NY, United States, Jonathan V. Hall, Daniel T. Knoepfle In many designed markets, the platform eliminates price dispersion by setting the product market price, yet still allows supply-side free entry and exit. We explore the equilibrium of such markets, using data from Uber. Following price increases,

drivers make more money per trip and—and initially more per hour-worked— and as a result, work more hours. However, this increase in hours-worked has a business stealing effect, with drivers spending a smaller fraction of hours-worked with paying customers, eventually bringing the hourly earnings rate back close to its previous level. The resulting higher fare/lower productivity equilibrium is generally inefficient. 4 - The Value of Time: A High-Frequency Analysis of Ride-Hail Auctions Nick Buchholz, Princeton University, NJ, United States We use detailed consumer choice data from a large ride-hailing application to study consumer valuations of time. This application offers a unique mechanism that allows taxi drivers in the Czech Republic to bid on trips, and allows consumers to choose between a set of characteristics of a ride: price, waiting time (based on the distance from consumers to each available taxi), car type and driver ratings. We observe the tradeoffs consumers face between prices and waiting times across 1.9M ride requests and 5M bids, as well as their ultimate selection. We leverage rich variation in bids and customer choices to directly measure consumer willingness-to-pay for time savings and decompose it further by time- of-day and location. Baseline estimates from Prague show the value of waiting to exceed median wages; from $0.40 per minute in peak weekday morning hours, to $0.05 per minute in off-peak weekend hours. Waiting times are important to explain customers’ choices, on average accounting for 13.7 percent of total valuation. We use our preference estimates to quantify the welfare benefits of offering a menu of time and price options. Our results show that a flexible menu of waiting time and choices improves welfare by up to 48 percent over a more standard dispatch mechanism. QSR Best Student Paper Award Sponsored: Quality, Statistics and Reliability Sponsored Session Chair: Youngjun Choe, University of Washington, Seattle, WA, 98105, United States Co-Chair: Matthew Plumlee, Northwestern University, Northwestern University, Evanston, IL, 60208, United States 1- Monitoring for Changes in the Nature of Stochastic Textured Surfaces Anh Tuan Bui, Northwestern University, Technological Institute 2145 Sheridan Road, Room C210, Evanston, IL, 60208, United States This work monitored for changes in the nature of stochastic textured surfaces without requiring prior knowledge of the changes. We used supervised learning to characterize the joint distribution of surface image pixels. Our monitoring statistic was based on likelihood-ratio principles. We demonstrated the approach with textile and simulation examples. 2 - Predictive Comparisons for Screening and Interpreting Inputs in Machine Learning Raquel Ferreira, Purdue University, West Lafayette, IN, 47907, United States A methodology based on predictive comparisons is specified to identify important inputs and interpret their associations with an outcome that are obtained via a machine learning algorithm. New predictive measures are constructed to perform input screening, and interpret conditional and two-way interactions of inputs. 3 - Scan B-Statistic for Kernel Change-Point Detection Shuang Li, Georgia Institute of Technology, Atlanta, GA, United States Detecting the emergence of an abrupt change-point is a classic problem in quality control, statistics and machine learning. In this paper, we propose a computationally efficient kernel-based nonparametric statistic for change-point detection, which enjoys fewer assumptions on the distributions than the parametric approaches and can handle high-dimensional data. A novel theoretical approximation of the tail probability of these statistics is proposed, which provides a convenient way to find the detection thresholds for both offline and online cases. We show that our methods perform well on both synthetic data and real data. 4- Scalable Robust Monitoring of Large-Scale Data Streams Ruizhi Zhang, Georgia Institute of Technology, Atlanta, GA, United States Online monitoring large-scale data streams has many important applications. However, it is non-trivial to develop efficient schemes due to three challenges: robustness, sparsity and computation. To tackle these challenges, we propose an efficient, robust and scalable monitoring scheme. Asymptotic analysis and extensive numerical simulations illustrate the usefulness of our proposed scheme. n SA68 West Bldg 105C

26