2016 INFORMS Annual Meeting Program

SD12

INFORMS Nashville – 2016

2 - Purchase Prediction Based On Multilevel Association Rule Mining Xinxue Qu, College of Business, Iowa State University, Ames, IA, 50010, United States, quxinxue@iastate.edu, Zhengrui Jiang Recommender systems are one of the most widely deployed applications in E- commerce. The goal of this study is to improve existing association-rule-based methods to increase the quality of product recommendations. There are two important factors in our method. First, due to the huge number of products in stores, market basket data is often sparse. Second, competing products are often highly substitutable, and consumers may be open to alternatives. The method we propose infers the level of similarity/substitutability between pairs of products from product category information. Experimental results show multilevel association rules can lead to a higher accuracy of purchase predictions. 3 - Risk Information Disclosure And Project Success Rate Yang Pan, University of Maryland, ypan@rhsmith.umd.edu Since information asymmetry between funders and creators is a critical issue in crowdfunding platform, many policies are introduced to improve the information transparency and make markets more efficient. One of the mechanisms is mandatory disclosure imposed by platform. We aim to understand how disclosed risk information has an effect on project outcomes. We study this question on a popular crowdfunding site requiring project creators disclose potential risk information about projects. We analyze the detail content of the disclosed risk information with text mining techniques and test the association between self- disclosed risk information and successful rate of crownfunding projects.

2 - From Local To Global Connections: A New Random Graph Model To Explain The Structural Properties Of Real-world Networks Rakesh Nagi, U of Illinois at Urbana-Champaign, Department of Industrial & Enterprise Systems, 117 Transportation Building, MC- 238, Urbana, IL, 61801, United States, nagi@illinois.edu Sushant Khopkar, Alexander Nikolaev Online Social Network (OSN) data are hard to interpret. Many OSN users have lots of connections, easily surpassing 150 - the Dunbar number. We present a random graph formation model that explains social tie formation by bridging the gap between the Watts-Strogatz and scale free networks. It shows how the information about “talented” individuals may propagate from their friends towards the masses, with a power law in degree emerging via the mechanism fundamentally different from preferential attachment (PA): while PA assumes full visibility, our model relies on local information exchanges. We report and interpret the model parameter estimates for several real-world networks. 3 - Constrained Sparse Optimization For Tensor Based Modeling Of Student Learning In Collaborative Environments Alireza Farasat, University at Buffalo, afarasat@buffalo.edu Educational systems have witnessed a substantial transition from traditional educational methods mainly using text books, lectures, etc. to newly developed systems which are artificial intelligent-based systems personally tailored to the learners. In this study, a constrained sparse tensor-based factorization approach is proposed for modeling of student learning in collaborative environments. The main challenge of modeling students learning is the fact that learning occurs over time therefor. We develop a probabilistic, constrained based approach to the tenser factorization model which enables capturing the underlying dynamics of students learning over time. 4 - Generalized Cascade Model And Seed Bounds For Disease Spread In Social Networks Arash Ghayoori, U of Illinois at Urbana-Champaign, Urbana, IL, 61801, United States, ghayoor2@illinois.edu, Rakesh Nagi In this talk, we introduce a new diffusion model for social networks, which generalizes most of the previously introduced diffusion models. We establish its relevance in disease spread (epidemiology) as well as viral marketing. An upper bound on the size of the influential set (“seed” set of nodes that if become infected, will eventually result in making the entire network becoming infected) is also obtained for a special case of this model. We show this bound to be tight by providing a simple algorithm that outputs an influential set with size nearly equal to this upper bound. SD12 104B-MCC Convexification Techniques in Integer Programming Sponsored: Optimization, Integer and Discrete Optimization Sponsored Session Chair: Sercan Yildiz, Carnegie Mellon University, 5000 Forbes Ave, Pittsburgh, PA, 15213, United States, syildiz@email.unc.edu 1 - Sparse Pseudoinverses Via LP And SDP Relaxations Of Moore-Penrose Jon Lee, University of Michigan, jonxlee@umich.edu Pseudoinverses are ubiquitous tools for handling over- and under-determined systems of equations. For computational efficiency and also in the context of identifying Gaussian models having a sparse precision matrix, sparse pseudoinverses are desirable. Recently, sparse left and right pseudoinverses were introduced, using 1-norm minimization and linear programming. We introduce new sparse pseudoinverses by developing tractable convex relaxations of the wellknown Moore-Penrose properties. In the end, we have several new sparse pseudoinverses that can be calculated via linear and semi-definite programming. 2 - Optimal Truss Topology Design By Mixed Integer Conic Optimization Tamas Terlaky, Lehigh University, terlaky@lehigh.edu, Mohammad Shahabsafa, Ali Mohammad-Nezhad, Luis F Zuluaga We present novel models, including Mixed Integer Linear Optimization (MILO) and Mixed Integer Second Order Cone Optimization (MISOCO) models, for Truss Topology Design Optimization. We discuss how classes of non-convex models can be reformulated as MILO and MISOCO models. We present our approach to solve the MISOCO models through adding Disjunctive Conic Cuts in a BCC framework. Additionally, we present an efficient line search method developed to solve the original non-convex model. Preliminary computational results indicate the effectiveness of our novel approaches.

SD10 103C-MCC INFORMS Prize Invited: INFORMS Prize Invited Session

Chair: Julia Morrison, Marriott International, Department 51/974.18, Bethesda, MD, 20817, United States, julia.e.morrison@marriott.com 1 - 2016 INFORMS Prize Presentation by GeneralMotors Michael Harbaugh, General Motors, Warren, MI, United States, michael.harbaugh@gm.com, Robert Inman, Peiling Wu-Smith, Yilu Zhang General Motors, 2016 INFORMS Prize Winner, will survey its sustained application of analytics and operations research. Highlights will include Vehicle Health Management: using advanced analytics to predict failure of certain automotive systems before customers are affected, Optimizing New Vehicle Inventory: determining first how much, and second what mix of vehicles to hold in dealer inventory, andRevenue Management for Vehicle Content and Packaging: leveraging customer preferences to package and price vehicle content that will sell best.

SD11 104A-MCC Network Optimization Sponsored: Optimization, Network Optimization Sponsored Session

Chair: Alexander Nikolaev, Assistant Professor, University at Buffalo, 312 Bell Hall, Buffalo, NY, 14260, United States, anikolae@buffalo.edu 1 - Optimal Seed Activation Scheduling For Influence Maximization In Social Networks Mohammadreza Samadi, Operations Research Consultant, American Airlines, Fort Worth, TX, United States, Mohammadreza.Samadi@aa.com, Alexander Nikolaev, Rakesh Nagi Influence maximization problem selects a set of influential nodes, called seeds, in a social network to spread the influence over the network maximally. We critique the basic assumption of influence maximization problem in the literature on controlling cascades only through the early starters and present Seed Activation Scheduling Problem (SASP) in two-level networks. The SASP is a sequential seed selection problem that results in optimal budget allocation over the campaign time horizon. The problem is modeled as a mixed-integer program for blogger- centric marketing campaigns and an efficient heuristic algorithm is presented using column generation method.

97