INFORMS 2021 Program Book

INFORMS Anaheim 2021

TD29

(categorical). For better interpretability and accuracy, we further introduce an attention layer that considers the criticalness of different events and a regularizer based on the time between consecutive events. We conduct numerical studies on simulated data and Computerized Tomography log data to evaluate our method. 2 - Active Sequential Change-point Detection under Sampling Control Qunzhi Xu, Georgia Tech, Atlanta, GA, United States This paper considers the active monitoring of multiple data streams for changes under the sampling control constraint. Here the sampling control constraint means that we are allowed to access only one local stream per time step. Under the scenario when the post-change distributions involve unknown parameters, we develop an efficient active sequential change-point detection algorithm: the greedy-cyclic-sampling-cumulative-sum (GCS-CUSUM) algorithm. It is surprising that our proposed GCS-CUSUM algorithm is asymptotically optimal to minimize the detection delay up to o(log ) subject to the average run length to false alarm constraint of when the dimension M= o(log( )) and goes to ∞ . Simulation studies are then conducted to illustrate the performance of the proposed algorithm. 3 - A Bayesian Deep Learning Framework for Interval Estimation of Remaining Useful Life in Complex Systems by Incorporating General Degradation Characteristics Minhee Kim, University of Wisconsin-Madison, 1513 University Ave, Madison, WI, 53706-1539, United States Abstract is not available TD31 CC Room 208A In Person: Platforms, Data, and Algorithms General Session Chair: Ali Makhdoumi, Duke University, Durham, NC, 27708-9972, United States 1 - Misinformation: Strategic Sharing, Homophily, and Endogenous Echo Chambers James Siderius, Massachusetts Institute of Technology, Cambridge, MA, 02139, United States, Daron Acemoglu, Asuman Ozdaglar We present a model of online content sharing where agents can “fact-check” to determine if this content contains misinformation. While agents value shares, they simultaneously fear getting caught sharing misinformation. With little homophily in the social network, misinformation is often quickly identified and brought to an end. However, when homophily is strong, whereby agents anticipate that only those with similar beliefs will view the article, misinformation spreads more rapidly because of echo chambers. We use this to show that a social media platform that wants to maximize engagement should propagate extreme articles amongst extremist users. This creates an endogenous echo chamber, or “filter bubble,” that is highly conducive to viral misinformation. We conclude with a few policy suggestions to combat platform incentives to recommend misinformation. 2 - Passing Data Directly into Prescriptive Analytics Lennart Baardman, University of Michigan, Ann Arbor, MI, 48103, United States Analytics has seen an increase in use to solve operational problems. Often, data- driven algorithms take a two-stage approach involving predictive and prescriptive analytics. Predictive analytics is used to develop models of uncertain quantities, which can then be used in mathematical optimization models of prescriptive analytics. In this work, we propose a framework that can directly translate data into decisions using only a single mathematical optimization model. Using a single model avoids any errors due to overfitting on predictions in a sequential approach or the misspecification of a predictive model. Additionally, our approach can deal with highly non-linear objective functions. We show the strength of our model both in theory and practice, and solves complex problems quickly. 3 - Optimal Data Acquisition with Privacy-Aware Agents Juba Ziani, Georgia Institute of Technology, Atlanta, GA, United States, Rachel Cummings, Hadi Elzayn, Vasilis Gkatzelis, Emmanouil Pountourakis We look at a data analyst that must optimally buy data from individually rational, privacy-aware agents, to compute a statistic. Agents derive utility from the output of the statistic, and trade-off the privacy costs they incur from revealing their data with the utility they get from the statistic. They may decide to opt out if their privacy costs are high. The analyst provides differential privacy guarantees to her computation, and adjusts the level of noise she offers to affect the agents’ privacy costs. She does so to optimize the accuracy of her computation, and must take into account the trade-off between i) adding more noise to incentivize participation and data collection through lower privacy costs and ii) adding less noise to obtain more accurate data points. We provide near-optimal algorithms for the optimization and mechanism design problems faced by the analyst.

TD29 CC Room 207C In Person: Applied Machine Learning for a Two-Sided Marketplace General Session Chair: Nathaniel Burbank, 1 - Algorithmic Contact Management in Sales & Service Luke Winslow, Wayfair, Boston, MA, United States, Akritee Shrestha

Contact center operations must balance demand and supply to achieve target service levels. Lack of an integrated framework for forecasting, traffic balancing, and routing decreases overall efficiency. In this presentation, we outline a three- level approach for contact optimization: 1. improving contact forecasting through machine-learning to enable advance scheduling, 2. shaping traffic in real-time to encourage or defer contacts based on existing staffing levels, and 3. routing contacts to the right team/agent based on predicted customer needs to maximize outcomes. This integrated framework helps us increase overall efficiency and improve service levels and outcomes on sales and service contacts. 2 - Bayesian Product Ranking for Multiple Objectives Tom Croonenborghs, Wayfair, Boston, MA, 02111, United States One of the most important aspects of a successful e-commerce business, especially one with the scale and breadth of products like Wayfair, is to make it easy for customers to find the perfect product for their need, right when they need it. Our catalog size presents a significant curation challenge: we need to balance exposing popular products with surfacing newer products that we believe can be successful, but have not yet received significant customer traffic. To this end, we have developed and deployed a Bayesian recommender system which not only finds opportunities for all of our customers, but at the same time maximizes benefits for our suppliers and Wayfair. 3 - Power Up Geo Experimentation at Wayfair With Integer Optimization Chenhao Du, Wayfair, Wayfair, Boston, MA, 02111, United States In Wayfair we run numerous experiments to evaluate the ad efficiency, optimize the UI interface and understand users’ preference. Comparing to the most popular user-level A/B experiments, Geo experiment is more privacy robust and able to handle the situations when user-level data is unavailable. In this talk, we will introduce how we improve the design and measurement on Geo experiment at Wayfair using the integer optimization. 4 - Share Of Voice: Optimizing Wayfair’s Marketing Content by Maximizing Customer Relevancy with Business Constraints Kurt Zimmer, Wayfair, 4 Copley Place, Boston, MA, 02111, Boston, MA, 02111, United States While Wayfair’s various ML models and pipelines provide us an algorithmic way to find the most relevant content for a given customer, there may be instances where we want to deviate from the most optimal short term strategy. In particular, this is often the case with new and emerging product categories, where we’d like to increase the Share of Voice of these categories by finding who are the most optimal customers to show these categories to. This is encoded as a Generalized Assignment Problem where we are aiming to assign each customer to a product category optimally given our constraints. We set the optimization objective to maximize the relevancy of the category to each customer as determined by our Customer Need Models and solve after adding the Share of Voice constraints from the business. Tests have shown a lift in engagement across both traditional and emerging categories. TD30 CC Room 207D In Person: Adaptive Online Learning of High-Dimensional Data General Session Chair: Mostafa Reisi Gahrooei, University of Florida, Gainesville, FL, 32608-1047, United States 1 - Deep Learning-based Critical Event Prediction using Time-dependent Representations Ye Kwon Huh, University of Wisconsin-Madison, Madison, WI, United States, Minhee Kim, Kaibo Liu In reliability analysis, event sequence data are commonly used in system monitoring, diagnosis, and critical event prediction. While many deep learning- based event prediction models have been proposed, they rely on representations (e.g., one-hot encoding) that do not fully incorporate the temporal information contained in the event sequence. To overcome this limitation, this study proposes a novel time-dependent representation of the event sequence by generating event embeddings that leverage both occurrence time (continuous) and event type

134