2016 INFORMS Annual Meeting Program

266 / 561

INFORMS Nashville – 2016
266
TB06
102A-MCC
Solving Hard Optimization Problems in
Machine Learning
Sponsored: Data Mining
Sponsored Session
Chair: Yan Xu, SAS Institute, Inc., 100 SAS Campus Drive, Cary, NC,
27513-2414, United States,
1 - Strategies For Maintaining Sparse Dual Solutions In Large-scale
Nonlinear SVM
Alireza Yektamaram, Lehigh University, Bethlehem, PA,
Joshua Griffin
This talk will focus on distributed methods for solving large-scale nonlinear SVM
problems. It is easy to show that the global solution for such problems may be
dense, resulting in large impractical models returned to the user. Further, the
accuracy of such models typically is poor compared to sparser approximate
solutions. This talk focuses on practical methods that can be used to directly and
efficiently seek out accurate sparse solutions regardless of whether or not the
global solution is dense.
2 - A Hessian Free Method With Warm-starts For Deep
Learning Problems
Wenwen Zhou, SAS Institute Inc, 100 SAS campus drive, Cary,
NC, 27513, United States, , Joshua Griffin
This talk will focus on solving deep learning problems with Krylov-based iterative
methods where effective preconditioning matrices are unavailable. For such
problems, convergence of the outer iterations can degrade when the iterative
solver repeatedly exits on maximum Hessian-vector products rather than relative
residual error . To address this issue, a new warm start strategy is proposed to
accelerate an existing modified conjugate gradient approach while maintain
important convergence properties. Numerical experience and addition to
convergence results will be provided.
3 - An Accelerated Power Method For The Best Rank-1
Approximation To A Matrix
Jun Liu, SAS Institute Inc., Cary, NC, United States,
, Ruiwen Zhang, Yan Xu
The best rank-1 approximation to a matrix is a fundamental tool in linear algebra
and many machine learning applications. The power method is one of the most
well-known approaches for computing the best rank-1 approximation to a
matrix. In this paper, we propose to accelerate the power method for the best
rank-1 approximation of a matrix by imposing an additional refinement step. The
refinement step combines the current approximate solution and the previous one
to obtain an optimally refined point that is used as an input to the next step of the
power method. Empirical results on synthetic and real data sets demonstrate the
effectiveness of the proposed method.
4 - Local Search Optimization For Hyper-parameter Tuning
Yan Xu, SAS Institute, Inc.,
Many machine learning algorithms are sensitive to their hyper-parameter
settings. In this talk we discuss the use of black-box local search optimization
(LSO) for machine learning hyper-parameter tuning. Viewed as a black-box
objective function of hyper-parameters, machine learning algorithms create a
difficult class of optimization problems. The corresponding objective functions
involved tend to be nonsmooth, discontinuous, unpredictably computationally
expensive. We apply a parallel hybrid derivative-free optimization algorithm that
can make progress despite these difficulties providing significantly improved
results over default settings with minimal user interaction.
TB07
102B-MCC
Networks and Data Analytics in Finance
Sponsored: Data Mining
Sponsored Session
Chair: Shawn Mankad, Cornell University, 401H Sage Hall, Ithaca, NY,
14853-6201, United States,
1 - A System-wide Approach To Measure Connectivity In The
Financial Sector
Sumanta Basu, Cornell University, ,
Sreyoshi Das, George Michailidis, Amiyatosh Purnanandam
We develop and estimate a system-wide measure of network connectivity for a
sample of very large financial institutions of the U.S. Our approach is in sharp
contrast with extant measures of systemic risk that, either explicitly or implicitly,
estimate such connections using pair-wise relationships between institutions. We
show that such a pair-wise approach may result in improper classification of
banks as systemically important. Our system-wide approach, based on a recently
developed Lasso penalized Vector Auto-Regression (LVAR) model, allows us to
detect important systemic events and identify systemically important institutions
in a statistically principled manner.
2 - The Topology Of Overlapping Portfolio Networks
Andreea Minca, Cornell University,
This paper analyzes the topology of the network of common asset holdings,
where nodes represent managed portfolios and edge weights capture the impact
of liquidations. We consider the degree centrality as the degree in the subnetwork
of weak links, where weak links are those that lead to significant liquidations. We
show that the degree centrality is correlated with excess returns, and is significant
after we control for the Carhart four factors.The network of weak links has a scale
free structure, similar to financial networks of balance sheet exposures. Moreover,
a small number of clusters, densely linked, concentrate a significant proportion of
the portfolios.
3 - Network Concentration And Systemic Losses
Agostino Capponi, Columbia University,
We develop a majorization-based tool to compare financial networks with a focus
on the implications of liability concentration. Specifically, we quantify liability
concentration by applying the majorization order to the liability matrix that
captures the interconnectedness of banks in a financial network. We develop
notions of balancing and unbalancing networks to bring out the qualitatively
different implications of liability concentration on the system’s loss profile. An
empirical analysis of the network formed by the banking sectors of eight
representative European countries suggests that the system is either unbalancing
or close to it, persistently over time.
4 - The Multilayer Structure Of The Financial System
Dror Kenett, US Office of Financial Research,
We introduce a new multilayer map to identify, quantify, and understand
interconnections that can spread a stress event across the financial system. The
network map has three layers showing the flow of assets, short-term funding, and
collateral that circulate among market participants. As we move from one layer to
the next layer, risk is transformed and spreads. For example, a price shock to one
type of securities in the asset layer may move through the network to become a
funding risk in the funding layer, then emerge as a counterparty or credit risk in
the collateral layer. We also discuss data gaps that must be filled to map the full
scope of interconnections in a multilayer financial system.
TB08
103A-MCC
Business Model Innovation
Invited: Business Model Innovation
Invited Session
Chair: Serguei Netessine, INSEAD, Singapore, Singapore,
1 - Sustainable Distribution Models At The Bottom Of The Pyramid
Bhavani Shanker Uppari, INSEAD Business School,
Ioana Popescu, Serguei Netessine
Although several products are invented to improve the lives of poor (e.g., efficient
cook stoves, solar lights), they do not necessarily reach the poor because these
people live in villages which are located beyond most multinationals’ distribution
networks. Lack of infrastructure and illiteracy only aggravate this problem.
Therefore, several firms rely on door-to-door (D2D) distribution networks to sell
their products. We investigate the strategic issues that arise in D2D models, such
as when is it appropriate to create a proprietary D2D network or share it with a
partner, what type of partner (for-profit vs. for-impact) is suitable, and how to
align the incentives of partners in a shared D2D model.
2 - Pay-as-you-go Business Models In Developing Economies
Jose A Guajardo, University of California-Berkeley,
Pay-As-You-Go business models have become widely adopted for the diffusion of
off-grid energy products in developing economies. In this research we provide an
empirical analysis of central aspects of this type of business models.
3 - Business Model Innovation Feature Extraction And Application To
The Lean Startup Framework
Christophe Pennetier, Insead,
Using a new curated dataset with more than half a million startups, we use state-
of-the-art text mining and machine learning techniques to identify business
model innovations and study their effects on startups’ success.
TB06