Informs 2015

353 / 552

INFORMS Philadelphia – 2015
351
TD18
18-Franklin 8, Marriott
Recent Advances in First Order Methods for Large-
Scale Optimization
Cluster: Modeling and Methodologies in Big Data
Invited Session
Chair: Mingyi Hong, Iowa State University, 3015 Black Engineering,
Ames, IA, 50011, United States of America,
1 - On the Information-adaptive Variants of the Admm:
An Iteration Complexity Perspective
Shuzhong Zhang, Professor, University of Minnesota, Department
of Industrial and Systems Eng, Minneapolis, MN, 55455, United
States of America, , Xiang Gao, Bo Jiang
We present a suite of variants of the ADMM, where the trade-offs between the
required information on the objective and the computational complexity are
explicitly given. The new variants allow the method to be applicable on a much
broader class of problems where only noisy estimations of the gradient or the
function values are accessible, yet the flexibility is achieved without sacrificing the
computational complexity bounds.
2 - An Optimal Randomized Incremental Gradient Method
Guanghui Lan, University of Florida, Gainesville, FL,
United States of America,
We present a randomized incremental gradient method and show that this
algorithm possesses unimprovable rate of convergence for convex optimization.
We provide a natural game theoretic interpretation for this method as well as for
the related Nesterov’s optimal method. We also point out the situations when this
randomized algorithm can significantly outperform the deterministic optimal
method.
3 - On the Expected Convergence of Randomly Permuted ADMM
Ruoyu Sun, Stanford University, Menlo Park, CA, 94025, United
States of America, , Zhi-Quan Luo, Yinyu Ye
Recently, it has been shown that the direct extension of the alternating direction
method of multipliers (ADMM) to the multi-block case fails to converge when
solving a simple square system of linear equations. In this paper, however, we
prove that, if in each step one randomly and independently permutes the
updating order of any given number of blocks, the method will converge in
expectation for solving the square system of linear equations.
4 - Alternating Direction Method of Multipliers for Distributed Sparse
Principal Component Analysis
Davood Hajinezhad, Iowa State University, 62 B Schilletter
village, Ames, IA, 50010, United States of America,
Mingyi Hong
We propose distributed algorithms to perform sparse PCA. They are quite flexible,
in the sense that they are able to handle different forms of data partition (i.e.,
partition across rows or columns of the data matrix). Numerical experiments
based on both real and synthetic data sets, conducted on high performance
computing clusters, demonstrate the effectiveness of our approaches.
TD19
19-Franklin 9, Marriott
Network Inference
Sponsor: Computing Society
Sponsored Session
Chair: Nedialko Dimitrov, Assistant Professor, UT Austin, University of
Texas at Austin, Austin, United States of America,
1 - Fast, Approximate Inference on Graphical Models by
Reducing Treewidth
Areesh Mittal, University of Texas at Austin, 1626 West 6th St.
Apt. F, Austin, TX, 78703, United States of America,
, Nedialko Dimitrov
Complexity of exact inference algorithms in graphical models is exponential in
treewidth. We develop technique to perform approximate inference by removing
edges and updating factors, leading to reduced treewidth. We prove bounds on
error in approximation. Finding updated factors involves solving a geometric
program (GP) with exponential number of constraints. We develop row
generation technique to solve the GP. We demonstrate the results on discrete
graphical models applied to social networks.
2 - Non-aggressive Adaptive Traffic Routing
Madhushini Narayana Prasad, Graduate Research Assistant,
Cockrell School of Engineering, University of Texas at Austin,
Austin, TX, 78712, United States of America,
, Nedialko Dimitrov
Routing a person through a traffic network presents a dilemma to choose
between fixed route which is an easier to navigate route and adaptive route
which minimizes the travel time by adjusting to the traffic conditions. We
investigate methods for non-aggressive, adaptive routing that is middle-ground
seeking the best of both these extremes, i.e. adaptive routes restricted in number
of route shifts allowed at a critical juncture, and investigate the trade-offs
between the extremes.
3 - Social Network Echo Chambers and Popularity
Yinhan Liu, University of Texas Austin, 1901 Crossing Place
#3301, Austin, TX, 78741, United States of America,
Nedialko Dimitrov
Social network users often have the goal of building a large follower base. Some
users are members of what we term echo chambers, a small group of users that
re-share each other’s messages. We present an empirical study on the impact of
echo chambers on the popularity of users using historical data from Twitter.
Specific questions we address are: Does echo chamber membership increase re-
shares outside the echo chamber? Does echo chamber membership increase
follower base?
TD20
20-Franklin 10, Marriott
Banking and Insurance
Contributed Session
Chair: Linna Du, Data Scientist, CACS, 2259 Adam Clyton Powell,
New York, NY, 10027, United States of America,
1 - Success Drivers of Online Equity Crowdfunding Campaigns for
Unaccredited Investors
Anna Lukkarinen, Aalto University, P.O. Box 21220, Helsinki,
00076, Finland, Jeffrey Teich,
Hannele Wallenius, Jyrki Wallenius
Using data from a leading equity crowdfunding platform in Northern Europe, we
explore success factors of campaigns. The results suggest that the investment
criteria traditionally used by professional investors are not of prime importance
for success in equity crowdfunding. Instead, success is related to pre-selected
crowdfunding campaign characteristics and networks. Understanding the success
factors of online equity crowdfunding campaigns is important to the design of
online platforms.
2 - The Mover-Stayer Process for the Credit Data
Anna Matuszyk, Assistant Professor, Warsaw School of
Economics, Niepodleglosci 162, Warsaw, 02-554, Poland,
Halina Frydman
Using the credit data set, coming from the European bank, we estimate the
mover-stayer model, which is an extension of the Markov chain. This model
assumes that the population is heterogeneous: there are “stayers” and “movers”.
“Movers” evolve according to a Markov Chain with the one-step transition
matrix, while “stayers” never leave their initial states. The probability of a
customer being a stayer in a paid up state is modeled using the logistic regression.
3 - Monopolistic Dealer Versus Broker: Impact of Proprietary
Trading with Transaction Fees
Yuan Tian, Ryukoku University, 67 Tsukamoto-cho, Fukakusa,
Fushimi-ku, Kyoto, Japan,
Katsumasa Nishide
We consider a one-period financial market with a monopolistic dealer/broker and
an infinite number of investors. While the dealer (with proprietary trading)
simultaneously sets both the transaction fee and the asset price, the broker (with
no proprietary trading) sets only the transaction fee, given that the price is
determined according to the market-clearing condition among investors. We
effectively demonstrate how proprietary trading affects market equilibrium and
welfare of investors.
4 - A Data Analytics Based Approach for Building 360 View of
Banking Customer
Tianzhi Zhao, IBM, Diamond Bld, ZGC Software Park, Beijing,
China, , Zhen Huang, Ming Xie, Bing Shao,
Yuhang Liu, Jian Xu, Wenjun Yin, Yuhui Fu
Banks today are experiencing transformation from product centricity to customer
centricity. With advent of big data, it enables banks to fully and deeply
understand customers by building 360 degree customer view. In this paper, a data
analytics based approach for 360 degree view of banking customer is proposed. It
can help banks quick build customer centricity for customer segments, targeted
marketing, personalized recommend, etc.
TD20