Content uploaded by Ahmed Abbasi
Author content
All content in this area was uploaded by Ahmed Abbasi on Nov 21, 2018
Content may be subject to copyright.
Advanced Customer Analytics: Strategic Value
Through Integration of Relationship-Oriented
Big Data
BRENT KITCHENS, DAVID DOBOLYI, JINGJING LI, AND
AHMED ABBASI
BRENT KITCHENS (bmk2a@comm.virginia.edu; corresponding author) is an assistant
professor of information technology (IT) in the McIntire School of Commerce at the
University of Virginia. He holds a Ph.D. in information systems from the Warrington
College of Business, University of Florida. His research interests include customer
analytics, health IT, and online information dissemination. He has published in various
journals, including Information Systems Research and Decision Support Systems.
Before pursuing a career in academia, he worked for five years in IT Risk Advisory at
Ernst & Young LLP.
DAVID DOBOLYI (dd2es@comm.virginia.edu) is a research scientist in the Center for
Business Analytics at the McIntire School of Commerce, University of Virginia. He
received his Ph.D. from the University of Virginia, and his primary research interests
involve quantitative modeling, experimental cognitive psychology, and artificial
intelligence, with applications including cybercrime and health. He has published
in numerous journals, including Science, and his publications span a broad range of
topics among which are reproducibility in science, eyewitness memory, Parkinson’s
disease, and learning styles theory.
JINGJING LI(jl9rf@comm.virginia.edu) is an assistant professor of information technol-
ogy in the McIntire School of Commerce at the University of Virginia. She received
her Ph.D. from the Leeds School of Business, the University of Colorado at Boulder.
Her research interests relate to machine learning and big data analytics, with applica-
tions in e-commerce, platform business, health care, search engine, user-generated
content, and recommender systems. She received the AWS Research Grant, and
Microsoft Research Azure Award for her work on big data analytics. Previously, she
was a scientist at Microsoft, where she proposed and implemented large-scale machine
learning solutions for Microsoft products such as Xbox One, Windows 8 Search
Charm, Windows Phone App Store, Cortana,andBing Entity Search.
AHMED ABBASI (abbasi@comm.virginia.edu) is Murray Research Professor of
Information Technology and director of the Center for Business Analytics in the
McIntire School of Commerce at the University of Virginia. He earned his Ph.D.
from the University of Arizona. His research interests relate to predictive analytics,
with applications in online fraud and security, text mining, health, and customer
analytics. He has published over 70 peer-reviewed articles in the leading journals
and conference proceedings, including Journal of Management Information Systems,
MIS Quarterly, ACM Transactions on Information Systems, IEEE Transactions on
Journal of Management Information Systems / 2018, Vol. 35, No. 2, pp. 540–574.
Copyright © Taylor & Francis Group, LLC
ISSN 0742–1222 (print) / ISSN 1557–928X (online)
DOI: https://doi.org/10.1080/07421222.2018.1451957
Knowledge and Data Engineering, and others. His projects on cyber security, health
analytics, and social media have been funded by the National Science Foundation.
He received the IBM Faculty Award, AWS Research Grant, and Microsoft Research
Azure Award for his work on big data. He serves as senior editor or associate editor
for several journals. His work has been featured in several media outlets.
ABSTRACT: As more firms adopt big data analytics to better understand their customers
and differentiate their offerings from competitors, it becomes increasingly difficult to
generate strategic value from isolated and unfocused ad hoc initiatives. To attain
sustainable competitive advantage from big data, firms must achieve agility in com-
bining rich data across the organization to deploy analytics that sense and respond to
customers in a dynamic environment. A key challenge in achieving this agility lies in
the identification, collection, and integration of data across functional silos both within
and outside the organization. Because it is infeasible to systematically integrate all
available data, managers need guidance in finding which data can provide valuable
and actionable insights about customers. Leveraging relationship marketing theory, we
develop a framework for identifying and evaluating various sources of big data in
order to create a value-justified data infrastructure that enables focused and agile
deployment of advanced customer analytics. Such analytics move beyond siloed
transactional customer analytics approaches of the past and incorporate a variety of
rich, relationship-oriented constructs to provide actionable and valuable insights. We
develop a customized kernel-based learning method to take advantage of these rich
constructs and instantiate the framework in a novel prototype system that accurately
predicts a variety of customer behaviors in a challenging environment, demonstrating
the framework’s ability to drive significant value.
KEY WORDS AND PHRASES: big data, customer acquisition, customer analytics, custo-
mer expansion, data integration, data management, design science, IT strategic
value, relationship marketing, customer retention.
As companies increasingly compete for customers’attention, raising the cost of
acquisition and increasing the difficulty of retention, the strategic challenge of under-
standing and managing customer relationships has become more difficult and more
important [26]. At the same time, it has become easier to obtain vast amounts of data
about customers that, when combined with increasingly sophisticated analytical tech-
niques, can provide important and actionable insights [40], allowing companies to
innovate and differentiate from competitors. However, with the proliferation of analy-
tics, much of the low-hanging fruit has already been claimed. As such, isolated, ad hoc
analytics initiatives can no longer achieve or sustain competitive advantage.
Companies that seek to attain strategic value through big data must create focused
analytics competencies that provide agility in adapting to a dynamic environment.
With the possible exception of a small number of young, technology-centric
companies built directly on analytics capabilities (e.g. Google, Amazon, Capital
One), customer analytics have largely been based on siloed data that capture a single
aspect of customer behavior at a time [18]. In the past decade, the proliferation and
advancement of web traffic analytics, online reviews, social media, customer
ADVANCED CUSTOMER ANALYTICS 541
relationship management (CRM) software, and other information technology-
enabled apparatus have caused a rapid expansion in the volume, variety, and velocity
of data about customers and their relationships with firms [5,13]. The firms that will
derive the most value from this expansion are those that progress from traditional
siloed customer analytics to what we designate as “advanced customer analytics,”
which integrate a rich variety of relationship-oriented data that enable a deep under-
standing of customers, driving actionable insights and outcomes for acquisition,
retention, expansion, and customer equity.
The integration and use of such rich data sources have great potential for driving
competitive differentiation and strategic value; however, there are substantial obsta-
cles to address. Data sources are often unstructured, noisy, and difficult to integrate
into a focused and cohesive view of the customer, leaving vanishingly few firms
with a coveted “360° view”of customers [34]. Overwhelmed by available and
potentially available data, companies must answer many questions, which often
fall to information technology (IT) managers [40,43]: “What data should we
collect?”“How do we prioritize data integration efforts?”“How do we quantify
the value provided?”and so on. Answers of “just collect and integrate it all”and
“we’ll figure out the value later”become infeasible as the quantity and variety of
available data outpaces the ability to collect, store, and process these data; therefore
these decisions must be increasingly informed and intentional [40]. While there have
been many calls for guidance in valuing data to inform such decisions [35,66], there
has been little research on the subject. With IT situated as a critical enabler for
advanced customer analytics, IT managers are in need of such guidance [40].
The current best practice solution to this problem of data collection and integration is
the creation of data lakes: vast repositories where all sorts of data from across an
organization are stored in their native format, just waiting to be analyzed and have
their value extracted by someone. Data lakes remove barriers and up-front costs for data
sharing across an organization, supporting experimentation and discovery. However, the
lack of standardization or integration inherent to this approach presents novel challenges
[27]. Whenever a new analytics initiative is introduced, data from the data lake must be
reorganized from scratch, including the daunting task of fishing relevant data out of the
lake and integrating it into a comprehensive view of the customer. More important, if
data are not intentionally collected in a way that allows linkages across data sources
(e.g., clickstream data with no link to customer IDs), integration may be impossible to
achieve with existing data, either stifling the analytics initiative through a lack of
relevant data or requiring time-consuming and costly one-off investments in a new
data infrastructure. Because this ad hoc approach to data integration and customer
analytics requires considerable up-front investment for each new analytics effort, it
impairs agility in sensing and responding to customer needs.
In short, a primary challenge in deriving strategic value from big data is the
difficulty of creating an integrated big data infrastructure that supports the agile
development of advanced customer analytics without overinvestment in worthless
data or underinvestment in data that could add significant value. To address this
challenge, we follow the design science paradigm to propose a novel framework and
542 KITCHENS, DOBOLYI, LI, AND ABBASI
system at the intersection of big data analytics, marketing, and IT strategy. Our
framework enables the development of advanced customer analytics systems that
harness the volume, variety, and velocity of available customer data while assessing
the value of various data sources as well as providing much needed guidance for
data management and integration decisions and investments.
Based on design principles from relationship marketing theory (RMT), our frame-
work consists of (1) a rich and generalizable set of relationship-oriented constructs
that provide insight into customer behaviors; (2) a principled, flexible, versatile
predictive model to extract value from a wide variety of structured and unstructured
relationship-oriented data; and (3) an approach for estimating the contribution of
various constructs for prioritizing data management and integration efforts.
Collectively, our framework is focused on building a data infrastructure for agile
deployment of customer analytics that leverage and improve customer relationships
to drive competitive advantage across a broad range of customer analytics use cases
where reliance on siloed data has impeded insight and business value.
We develop a novel kernel-based machine learning method to serve as the
predictive model in our framework. This method combines a radial basis function
kernel with novel hybrid tree and weighted cross entropy string kernels in a
composite kernel support vector machine (SVM). This innovative method allows
for the principled embedding of theory and domain knowledge into the constituent
kernels; is flexible in incorporating a wide variety of data in tabular, graphical, and
text format; and provides versatile ensemble-like performance across a wide portfo-
lio of customer analytics tasks.
We instantiate the framework in a novel advanced customer analytics prototype
system developed for a major U.S. e-commerce and catalog-based retailer of educational
materials. We evaluate the system and underlying framework using 664,737 actual
customers sampled from the firm. The prototype system implements a portfolio of
customer analytics applications, accurately predicting customer churn, conversion on
specific promotional offers, and lifetime value, all within 30 days of first purchase.
Deployment of this portfolio results in significant value for the firm. We also assess the
value of various potential data constructs relative to data management and integration
costs, offering guidance and justification for investment in data infrastructure that
provides agility for deploying further analytics. Because of our work, our corporate
partner made significant investments in expansion and integration of data sources found
to be most valuable by our prototype, in order to support analytics initiatives.
Our research makes several academic and managerial contributions. Our primary
contribution is the synergistic ecosystem of closely related design science artifacts
that enable the creation of a value-justified infrastructure for the rapid and agile
deployment of a portfolio of advanced customer analytics. Our proposed framework
supports development of advanced customer analytics capabilities, directly addres-
sing the challenges of determining which data to invest in and integrate. The novel
composite kernel SVM we develop supplies a method tailor-made for extracting
insight and value from a rich variety of structured and unstructured relationship-
oriented data. The instantiation of our prototype system demonstrates that our
ADVANCED CUSTOMER ANALYTICS 543
framework can be implemented in a working system, providing significant value in a
complex real-world environment and demonstrating “last mile”relevance [47].
These artifacts form a valuable contribution to the literature in design science and
strategic value of big data.
Our research also has managerial implications for marketers performing customer
analytics and IT managers who are asked to support big data analytics initiatives. We
provide a general framework for designing advanced customer analytics that keeps both
value and costs in mind. As IT managers are asked to prioritize and justify the value of
data management and integration activities to support big data analytics [40], the ability
to measure the value of each data channel becomes an important consideration [11,24].
Our broadly applicable framework identifies a variety of novel constructs and provides
guidelines for measuring the value of each, supporting data management decisions.
Finally, our research contributes to the literature regarding the use of big data for
predictive analytics [5,13,22]. Recently, there have been many calls for research
regarding the use of big data for decision making to determine whether and how it adds
value to organizations. Shmueli and Koppius [56] note a dearth of research examining the
value of predictive analytics. Goes [22] specifically extolls the potential benefits of
combining a variety of “micro data”from different sources that are enabled by big data
analytics. Agarwal and Dhar [5] suggest researchers investigate the resultant ability to
study not only organizational and societal but also individual micro-level outcomes.
Abbasi et al. [2] point out the need for research that quantifies the value of the volume
and variety of data used in big data analytics relative to costs in order to demonstrate
effectiveness of big data investments and evaluate the feasibility and efficacy of big data
IT artifacts. The current study adds to this nascent literature on big data analytics for
prediction, with particular attention to the value of analytics for strategic decision making.
Literature Review
IT Strategy and Support of Big Data Customer Analytics
In order to create sustainable strategic value in a competitive landscape characterized
by rapid change, firms must build dynamic capabilities for adapting, integrating, and
reconfiguring resources to match their environment [31]. These dynamic capabilities
provide strategic agility, which allows firms to quickly recognize and capitalize on
opportunities, thereby creating competitive advantage. Sambamurthy et al. [54]
demonstrate that IT competencies, when effectively integrated with overall business
strategy and capabilities, serve as a platform for agility. However, IT investments to
improve agility must be carefully planned and managed, as haphazard or misguided
investment can actually impede agility [38].
A burgeoning opportunity for IT to support competitive action exists in the field of
big data analytics [12]. In the current environment and for the foreseeable future,
analytics represents a primary arena for innovation and competition. The prolifera-
tion of abilities to harness big data to improve decisions, processes, and products has
made such capabilities a basic requirement for survival, with those best equipped to
544 KITCHENS, DOBOLYI, LI, AND ABBASI
extract value from data achieving significant competitive advantage [18]. To derive
strategic value from analytics, it is important that firms innovate by: (1) moving
from general-purpose to specialized analytics uniquely optimized to address specific
business issues; and (2) eliminating organizational silos to coordinate data sharing
and analytics across functional boundaries [50].
Achieving these two objectives is prohibitively expensive and time-consuming
through ad hoc efforts. Instead, IT must strategically partner with other business
functions and become a proactive advocate and architect for analytics [57].
Specifically, IT departments should provide data governance and infrastructure to
support agile integration of data from multiple sources, organizing “around data as
if it were a valuable organizational asset”[50, p. 15] to foster innovation and sustained
competitive advantage from big data analytics [66]. By creating an infrastructure that
incorporates data across functional silos, “resulting client services are superior, less
susceptible to commoditization, and generate higher revenue”[39, p. 217 (emphasis
added)]. If achieved, this structure can provide a foundation for establishing a portfolio
of specialized yet coordinated analytics initiatives that deliver strategic value.
A key challenge relates to the collection and integration of valuable data across
silos within and outside the organization. Data management has long been consid-
ered a cornerstone of the IT function [23], and “big data’s rise has further amplified
the importance of IT in this role”through challenges and opportunities of exponen-
tially increasing data volume, variety, and velocity [2, p. 2]. Together these aspects
bring into focus the need for data infrastructure investment as well as the potential
value of resulting analytics [5,13,22]. However, unfocused data management and
integration can be extremely costly, and benefits are not always sufficient to offset
these costs [24]. Research suggests that “no single integration strategy is optimal in
all cases”[11, p. 89], and an intermediate level of integration is often more
beneficial than complete integration [11,24,40]. In order to support big data
analytics through an integrated data infrastructure, organizations must find effective
strategies to assess the value of available data.
Given the wide variety of available sources of relevant data, customer analytics
initiatives stand to gain significantly from such value-driven investments in IT infra-
structure for integration. By building on one of the firm’s most important resources (its
customers) and one of its least imitable (its data), customer analytics represent an
important strategic initiative with the potential to create significant and sustainable
competitive advantage. Customer analytics are enabled by a firm’scustomeragility—
the ability to sense opportunities for innovation and respond to those opportunities
with competitive action—and operational agility—the ability to rapidly redesign
processes to exploit marketplace conditions [54]. These dynamic capabilities should
be supported by the synergistic combination of interfunctional business coordination
and IT infrastructure for integrating the right data across the organization [53].
As firms move toward analytics specialization, there are opportunities to create a
diverse portfolio of customer analytics initiatives spanning acquisition, retention,
and expansion in order to optimize customer lifetime value and equity [26]. In order
to be most effective, this portfolio of specialized initiatives should draw from various
ADVANCED CUSTOMER ANALYTICS 545
aspects of the organization to incorporate the data most valuable for accomplishing
each individual objective. For this to become feasible, a common framework is
needed for designing customer analytics applications that incorporates business and
IT strategy. With this, firms could build a value-justified infrastructure for supporting
a portfolio of advanced customer analytics capabilities that combine to create
significant strategic value and sustainable competitive advantage.
Customer Acquisition, Retention, and Expansion
Analyses across the customer lifecycle—encompassing acquisition, retention, and
expansion—have become a critical focus for firms as many shift from a historically
product-centric orientation to one that is more customer-centric, with emphasis on
retaining and building profitable relationships with customers [26]. Current approaches
to examining customer retention may largely be traced back to Schmittlein et al. [55],
who focus on simple patterns in customer purchase activity to estimate the probability
that a customer is still “alive”(i.e., will continue to repurchase in the future). Using
only the observed recency, frequency, and monetary value (RFM) of purchases, the
model provides probabilistic predictions for individual customer churn. This work
originated a stream of research that has had much success at generating parsimonious
models that are practical for accurately predicting behaviors for individual customers
[20]. However, customers must have substantial purchase histories for the models to
be effective, and thus the settings considered by this research are limited to businesses
with relatively high purchase volumes. Outside of these high-volume environments,
the models lose predictive power.
Aside from the RFM paradigm that dominates customer analytics in marketing
literature, understanding customer behavior has become a common use case for the
application of machine learning, with many different techniques employed. Neslin
et al. [46] were among the first to describe machine learning for customer analytics
in a paper summarizing the results of 44 entries to a contest for predicting customer
retention, determining which methodological choices impact the quality of predic-
tion. Subsequent studies have applied various methods of all degrees of sophistica-
tion [16,36,46]. However, these studies focus on the machine learning methods
themselves, largely agnostic of the specific setting of customer analytics. Most
completely omit details of the setting and data set, with no description of how
customer behaviors were measured, what features were included, or how the results
could be used in context. These studies contribute to knowledge regarding model
selection aspects of customer analytics without regard to other inputs to the analytics
process. Of the studies that do provide details, many rely heavily on transactional
RFM constructs. For instance, Ballings and Van den Poel [6] employ demographic
and transactional features to predict churn. They show that reasonable model
performance with these features can require five or more years of data for training.
But given the importance of timeliness in the pursuit of strategic business value [13,
546 KITCHENS, DOBOLYI, LI, AND ABBASI
40], there is a strong need for relationship-oriented analytics that can produce results
in weeks rather than months or years.
Research Gaps
From our analysis of relevant prior literature, we identify three important
research gaps. First, there is a lack of research providing guidance in evaluating
the impact of data management capabilities at a granular level that could inform
prioritization decisions for focused data management and partial integration,
which has been called for in the literature [11,24]. While some studies have
focused on the strategic value of data management capabilities at the organiza-
tional level or benefits versus related costs of collecting and integrating certain
data for specific purposes [43], none provide a general framework for valuing
data for analytics initiatives in a broader sense. While there have been calls for
this type of research, little has been undertaken. This ability to value various data
sources is imperative for developing an integrated big data infrastructure for the
agile development of analytics.
Second, there is a lack of studies that have taken a comprehensive and holistic
relationship view, as opposed to a siloed transactional view, in predicting custo-
mer behaviors. As discussed, prior studies have largely focused on RFM [20,55]
or practically whatever data are available [36,46]. There has even been criticism
of this approach, heretofore unaddressed, pointing out that a vast majority of
customers at many firms have made only a single purchase and are consequently
largely indistinguishable from an RFM perspective, rendering the approach use-
less [41]. As we will discuss later, many studies have evaluated relationship-
oriented constructs associated with customer behaviors using explanatory models
[7,17], yet few have been utilized for prediction, and few employ a multifaceted
perspective, whether in a predictive or explanatory context. A more holistic and
integrated view of customer relationships would provide more accurate and
broadly applicable analytics.
Third, there is a paucity of literature on how to formally operationalize the
relationship-oriented constructs and data valuation appraisal alluded to in the first
two gaps for the purposes of performing predictive customer analytics. From a
design science perspective, instantiations offer essential prescriptive guidelines and
proofs-of-concept regarding how IT artifacts might be developed to solve an impor-
tant class of problems [28]. Such instantiations move beyond conceptual claims and
offer practical, research-grounded operationalizations that not only advance the
literature but also inform practice.
In this study, we address these gaps. We present a framework specifically focused
on providing a structure for determining the value of individual data sources, both
alone and in combination, in order to help IT managers decide where to invest
resources and how to justify data management and integration efforts in support of
advanced customer analytics. The framework includes the identification of a rich set
ADVANCED CUSTOMER ANALYTICS 547
of relationship-oriented constructs for use in advanced customer analytics systems.
Our prototype system and novel machine learning method, evaluated in the live
setting of an e-commerce retailer, demonstrates the significant value that analytics
developed using this framework and construct set may provide.
Framework for Advanced Customer Analytics
The design science research paradigm provides principles for the creation of inno-
vative IT artifacts that allow organizations to improve capabilities, address ever more
challenging issues, and take advantage of new opportunities, including those
afforded by big data [28,64]. We propose a framework for implementing advanced
customer analytics solutions, including a rich set of constructs that may be used to
describe and predict consumer behavior. We also create a novel kernel-based learn-
ing method to extract value from this rich variety of structured and unstructured data.
Using the framework and method, we instantiate a prototype customer analytics
system in a challenging environment requiring rich, relationship-oriented data to
make accurate predictions not possible using siloed, transactional data.
Under the design science paradigm, kernel theories are often used to inform the
development of novel design products [64]. In the design of our framework, method,
and instantiation, we primarily reference the theory of relationship marketing.
Relationship marketing theory (RMT) frames interactions with customers not as a
set of discrete transactions, but as a relational exchange. Relationships with custo-
mers provide a source of sustainable competitive advantage and strategic value for
the firm [48]. Combining customer perceptions, preferences, interactions, and com-
munication via various channels, this humanized view of customers stands in
contrast to the mechanical view of traditional customer analytics.
RMT has specifically informed a body of literature focused on understanding how
perceptual constructs such as satisfaction and repurchase intentions influence customer
behaviors through explanatory rather than predictive models. Crosby and Stephens [17]
provide some of the earliest research representative of this stream. Surveying a sample of
customers, they measure otherwise unobservable opinions and perceptions, including
various aspects and antecedents of satisfaction with the purchase (in this case, whole life
insurance). They show that more satisfied customers are more likely to renew or upgrade
their policies with the same company. Much research has followed this approach, using
surveys to measure perceptual constructs related to many aspects of customer relationships
with firms and assess how they influence future behavior. Beyond satisfaction, studies have
examined the impact of service failures, demographics, service usage, and attachment
styles on customer retention [7,14,44].
RMT provides design principles that guided all aspects of our proposed framework (see
online Appendix A for detail). The relationship marketing literature is extensive and varied
but has a distinct focus on understanding customer relationships as opposed to predicting
them. In addition to implementing modeling techniques geared toward explanation rather
than prediction, many studies include perceptual constructs operationalized through
548 KITCHENS, DOBOLYI, LI, AND ABBASI
surveys, which makes implementing the resultant models for individual-level predictions
difficult. However, the concepts introduced in this literature are informative in building
predictive analytics for nuanced problems requiring a holistic and integrated view of
customer relationships from a variety of sources. Figure 1 depicts the proposed framework,
including:
1. A rich set of relationship-oriented constructs to guide the identification and
acquisition of valuable data for advanced customer analytics
2. A principled, flexible, versatile predictive model for extracting value from
data constructs
3. Value-based evaluation metrics for action- and outcome-oriented cost/benefit
analysis
4. Construct evaluation leading to value-justified integration and data manage-
ment investments in IT infrastructure to create a foundation of integrated data
for analytics
5. A portfolio of advanced customer analytics capabilities supported by IT-
enabled agility for providing strategic value through enhanced customer
lifetime value and customer equity
Figure 1. Advanced Customer Analytics Framework
ADVANCED CUSTOMER ANALYTICS 549
Advanced Customer Analytics Constructs
In order to guide the identification and acquisition of valuable data sources, we
develop a comprehensive and generalizable set of potential construct categories
beneficial for advanced customer analytics. Many of these have rarely been used
to predict customer behavior due to the difficulty of integration. The synergistic
value of integrated data from a rich variety of sources is the key benefit of advanced
customer analytics. To identify the constructs for consideration in advanced custo-
mer analytics systems, we performed a broad search of the relationship marketing
literature.
1
Here we briefly describe each of the identified constructs and the
rationales for their inclusion.
Transaction: Transaction characteristics represent the most basic construct in
regard to customer relationships and form a basis for any customer analytics solu-
tion. As discussed, the most popular approaches for predicting customer behaviors
rely solely on measurements of the recency, frequency, and monetary value (RFM)
of customer transactions [20,55]. Transaction characteristics may include useful
details in addition to RFM, such as discount levels, promotions, payment types,
and specific items purchased, all of which may provide valuable information about
customer characteristics and behaviors.
Demographics (and Intangible Customer Characteristics): Demographics are also
among the most basic of characteristics firms can measure about their customers.
Relationship marketing studies have specifically found that demographic characteris-
tics influence customers’baseline expectations and thresholds for other constructs
such as satisfaction, influencing repurchase intentions and behavior [44]. Other studies
have specifically focused on how demographics impact customer behavior or simply
incorporate demographics as basic features in models focused on other data sources
[36,52,59]. The combination of transaction and demographic characteristics form
what we consider the basic starting point for designing a customer analytics solution.
Other intangible customer characteristics may influence relationships with a firm, such
as expertise, perceived risk/risk affinity, and technical self-efficacy [8].
Engagement: From an RMT perspective, customers may engage with firms in
many ways aside from purchase transactions, with the quantity and types of engage-
ments offering important clues regarding future behavior [19]. A customer may visit
a website, open an e-mail, call customer service, post a review or opinion, participate
in a forum, and so forth. Engagement provides a signal of a customer’s interest in a
continuing relationship with the firm. Relationship marketing research has often
focused on customer service interactions related to service failures and recoveries [7,
25]. Other studies include basic constructs related to engagement, such as usage
patterns, loyalty program and “wish list”usage, and e-mail opt-in/opt-out [62]. There
is much variability in the sophistication with which firms are able to collect,
measure, and analyze engagement interactions, particularly through outlets not
controlled by the firm (e.g., social media), which may have led to the dearth of
research investigating its value [19,63]. As these capabilities improve, customer
engagement has potential as a rich resource for advanced customer analytics. A key
550 KITCHENS, DOBOLYI, LI, AND ABBASI
challenge with engagement data is integrating it with other customer data for
analysis. Our framework provides a structure for identifying and valuing such
information, supporting integration efforts.
Satisfaction (and Related Perceptions): As discussed previously, customer satis-
faction is a key, foundational construct in RMT [7]. Many studies have shown that
increasing satisfaction leads to better customer retention and higher customer life-
time value [7,44]. Beyond satisfaction, other intrinsically related perceptual con-
structs studied include trust, commitment, loyalty, and other perceptions of firm and
environmental attributes [61]. The key challenge with satisfaction and related per-
ceptions is that they can be difficult to measure for individual customers. Most
studies examine these perceptions on a sample basis through surveys, with the intent
of explaining customer relationships or measuring entire market share rather than
making individual-level predictions [17,44].
Choice: A key observation available from a customer’s first as well as subsequent
purchases is the choice of product(s) purchased. Particularly when a firm sells goods
or services that are horizontally differentiated over differing customer preferences,
understanding the types of products a customer prefers can be useful in several ways.
From an RMT perspective, some have argued that the categories and types of products
purchased by a customer can be an indicator of their level of trust and interest in a
particular company [42]. The amount of variety in products purchased may also
provide a signal, with cross-buying increasing switching costs as consumers become
more aware of the firms offerings and quality [52]. Product choice can also be
informative regarding what products the consumer would be interested in purchasing
in the future, and product assortment is an important factor in retaining customers.
Channel: The channel(s) through which customers are acquired and continue to
interact with a firm provides information about customers and sets the stage for ongoing
relationships. RMT suggests that various channel interactions may impact the loyalty or
connection customers feel to the firm [8]. For instance, a customer calling and speaking
to a representative may develop a stronger relationship with the firm as a result of this
communication. Studies have also suggested that customers acquired through digital
channels tend to be more loyal and active due to self-selection effects and greater
opportunities to form connections with the company [29]. In addition, each channel is
likely to attract a different type of customer [32]: for example, online channels may
attract those who are more technology-savvy and hedonic.
Messaging: From an RMT perspective, the communications the customer receives
from the firm can also have a profound impact on future behavior [17]. Messaging can
often be focused specifically on selling: for instance, by using promotional offers to
entice a purchase. However, relationship-building messaging specifically focused on
enhancing the customer’s view of the firm is effective at retention, yet promotional
messaging, while effective in the short term, can have various effects over time [21].
The mode and quantity of communications received may also impact customer
behavior, with both too little and too much communication being detrimental [59].
Firm and Environment Characteristics: Characteristics of the firm and the market
environment in which it operates set the stage for the customer’srelationshipwiththe
ADVANCED CUSTOMER ANALYTICS 551
firm. Brand equity, payment equity/perceived fairness of the firm’s pricing policy, and
firm ethics and citizenship have all been shown to play a significant role in customer
perceptions and relationships with the firm [61]. Outside the firm, customer relation-
ships may be influenced by characteristics of the environment, such as dynamism,
munificence, complexity, competition and characteristics of competing firms, and
market share [8,62]. These represent important features for incorporation into analytics
applications, but also importantly should inform the entire analytics process.
Particularly, changes in characteristics or perceptions of the firm and environment
should be monitored in order to update analytics solutions for continued relevance.
Principled, Flexible, Versatile Predictive Model
Relationship-oriented data can be nuanced and take on many different formats, both
structured and unstructured. Yet, traditional machine learning methods are limited in
the complexity of data they can analyze. In order to ensure full extraction of
available value within each construct category for producing customer insights, the
framework should employ a predictive model that is:
●Capable of principled/theory-driven prediction to exploit the nuances of
relationship-oriented data
●Flexible in accommodating complexity of a wide variety of data constructs of
various formats
●Versatile in application to a variety of customer analytics initiatives
Kernel-based learning methods embody each of these characteristics, allowing
theory embedding via customized kernel functions, incorporation of structured and
unstructured data, and adaptability to a wide range of tasks [1,15], making it
possible to effectively represent the complexities of customer behavior.
Value-Based Evaluation Metrics
Aside from identification of a comprehensive set of constructs—the variety of which
provide a basis for performing advanced customer analytics—a key contribution of our
proposed framework is the presentation of guidelines for directly measuring the value of
an analytics solution, both in whole and relative to its constituent parts. This is a
necessity for advanced customer analytics endeavors, as essential support for manage-
ment of the volume, variety, and velocity of required data can be costly and must be
justified. Synthesizing the nascent literature on the value of big data analytics [40,51]
and infonomics [35], the key steps for creation of value-based evaluation metricsfollow.
Identify Potential Action(s): An intermediate outcome of the final advanced customer
analytics product will be predictions regarding customer behavior. But in isolation
predictions are useless: only when predictions inform decisions to take action do they
have the opportunity to create value for the organization [51]. These decisions and
actions enabled by the system are what ultimately provide actual value to the
552 KITCHENS, DOBOLYI, LI, AND ABBASI
organization through better relationships with customers, influence on customer pur-
chase behaviors, reduced costs, and so on. It is therefore critical that any advanced
customer analytics solution has these final value-driving decisions and actions as an
ultimate goal. Thus, the first step in measuring the value of the advanced customer
analytics solution is to determine potential actions to be taken once predictions are made.
In the case of churn prediction, one potential action might be to eliminate all marketing
efforts with positive marginal costs to customers predicted to have a high churn
propensity [52]. Alternatively, a targeted retention campaign might be created to attempt
to convert those with high churn propensity into returning customers [16,36]. Various
possibilities exist. The key is to strategically consider options before and during the
design of the advanced customer analytics application, incorporating the potential
actions into evaluation of the solution.
Determine Action Value: Once the potential actions to be taken on predictions are
known, a value must be determined for taking the proposed action on each customer.
This process is similar to constructing a cost matrix for cost-sensitive classification
models [1]. A cost or benefit must be assigned to correct and incorrect positive and
negative predictions. In the churn case, assuming the potential action is cessation of
marketing activities to likely churners, costs and benefits to be considered include:
(1) the marginal cost of marketing to a customer, which could be saved for those
customers predicted to churn; and (2) the expected revenue to be gained from a
customer who continues to make purchases, which may be lost for those who stop
receiving marketing materials [52]. Determining the values for potential actions will
require assumptions based in part on analysis of historical activities [1].
Define Cutoffs: With potential actions and related values determined, the final step
before implementing the value-based evaluation metric for measuring competing
models is to define which customers will receive action. This need not be the entire
set of customers, and in many cases, particularly with imbalanced classes, may
represent a small proportion of the whole [56]. In the example given previously,
marketing was to be discontinued to those customers with a high churn propensity.
Any model implemented in an advanced customer analytics solution will only
provide an estimate of the likelihood of a certain outcome for each customer.
Careful consideration must be given to how to determine what share of customers
will receive specific potential actions [52,60].
Generate Final Metric: The actions, value, and cutoffs determined in this step are
combined into a complete value-based evaluation metric, which is critical in later steps
to evaluate constructs and models, guide data management, integration, and real-time
feature engineering efforts, and demonstrate potential return on investment for market-
ing and IT managers to justify their efforts. For each combination of potential action
and actual outcome, a net cost/benefit is defined and applied to all predictions. In order
to arrive at a total expected value, the per customer value is multiplied by the number
of customers in the live population who exceed the cutoff. The value for an entire
analytics portfolio may be calculated by aggregating the total expected value for each
individual application. This sum may be used for evaluating and justifying invest-
ments in infrastructure to support advanced customer analytics capabilities.
ADVANCED CUSTOMER ANALYTICS 553
Construct Evaluation and Value-Justified Data Management,
Integration, and Feature Engineering
As mentioned, the cost of data management, integration, and real-time feature engineering
is among the most important of considerations in any big data analytics initiative [11,40].
A key aspect of the proposed advanced customer analytics framework relates to evaluating
each potential construct and data source with regard to the value-based metrics identified
earlier, allowing IT departments supporting advanced customer analytics initiatives to
focus efforts on integrating data from only the most valuable sources. In order to quantify
the individual value of each construct, we propose the use of an add-in/leave-out approach.
Comparisons may be made for each individual construct by running models excluding it
from the full model with all constructs and also by adding it to the base model with only the
most basic constructs available for production use without additional investments in data
management and integration. Additional models may be run with subsets of constructs
deemed to contribute significant value. The result of this analysis provides input for final
decisions regarding which constructs and data sources should receive investments in data
management, integration, and real-time feature construction. The value of each construct
and data source may be compared to the effort required in these areas to provide inputs for a
final production model and may be used to justify related expenditures and identify the
most valuable sources for partial integration [11,24]. This results in value-justified
investments in IT infrastructure supporting agile deployment of advanced customer
analytics.
Portfolio of Advanced Customer Analytics
The ultimate goal of the framework is to enable the design of a portfolio of advanced
customer analytics applications supported by this value-justified IT infrastructure and
the agility it enables. This portfolio should include applications ranging across custo-
mer acquisition, retention, and expansion. By enhancing customer relationships across
these dimensions, the portfolio can be used to improve customer lifetime value and
customer equity, creating strategic value and sustainable competitive advantage [26].
Prototype System Instantiation
Based on the proposed framework for advanced customer analytics, we develop a proto-
type system to show that the framework can feasibly be implemented in a working system
[28]. In order to demonstrate suitability to our framework’s intended purpose of supporting
the creation of advanced customer analytics, we must identify an application representative
of this circumstance. The defining feature of advanced customer analytics is the ability to
provide predictive insights into problems not solvable using a traditional siloed view, but
instead requiring the support of rich, relationship-oriented data integrated from a variety of
sources. Despite the fact that a significant amount of research has been produced regarding
customer behavior prediction, the transaction-oriented methods provided require long
554 KITCHENS, DOBOLYI, LI, AND ABBASI
customer histories (often five or more years) and are less accurate and effective outside
high-purchase-volume environments [6]. Many firms provide goods or services that are
purchased infrequently by nature or attract a large number of single-purchase customers,
and no existing method addresses this challenging case. We propose that advanced
customer analytics based on rich, relationship-oriented customer data will provide an
effective solution in this setting where other methods fall short.
To create our prototype system, we partnered with a company that we will refer to as
Course Shop International (CSI), a large e-commerce and catalog-based seller of educa-
tional materials for lifelong learners. CSI is highly representative of the high single-
purchase, low-frequency environment of interest. A large majority of CSI’scustomers
make only a single initial purchase, and of those who return, many months may pass
between purchases. The goal of our prototype system is to make predictions about
customers’future behaviors just 30 days after their initial purchase, when these predictions
are most valuable. During this period, we observe aspects related to each construct in our
framework. An illustration of the problem setting for our prototype system is provided in
Figure 2. To demonstrate how firms may use our framework to develop a data infra-
structure providing agility in supporting a portfolio of analytics initiatives, our prototype
system consists of three distinct customer analytics applications:
●A churn prediction application focused on evaluating which customers the
firm should invest in through continued marketing efforts (retention)
●A conversion prediction application for identifying customers likely to
respond to individual email promotions to reduce messaging fatigue and
prevent attrition (retention/expansion)
Figure 2. Problem Setting: Customer Agility through Advanced Customer Analytics
ADVANCED CUSTOMER ANALYTICS 555
●A customer lifetime value (CLV) prediction application for identifying custo-
mers who could successfully be expanded through offer of participation in a
premium loyalty program (expansion)
These three applications allow us to evaluate each aspect of our framework and the
strategic value it creates. Here we focus on the churn prediction example. Details of
Results of the combined portfolio are discussed in the evaluation section, while
details of the CLV and conversion prediction tasks are provided in online Appendix
B.
Leveraging the proposed advanced customer analytics framework described pre-
viously, we implemented a novel churn prediction system for CSI. Figure 3 shows
the system diagram, encompassing five stages: Data Lake, Feature Generation,
Data Preparation, Offline Modeling/Evaluation, and Online Implementation. These
five stages are closely aligned with facets of the proposed framework. For instance,
the Data Lake, Feature Generation, and Data Preparation components of the
system relate to the Investigate, Identify, and Acquire Potential Data section of the
framework. Offline Modeling/Evaluation is associated with Predictive Modeling and
Value-Based Evaluation Metrics. Lastly, the Online Implementation component
relates to the framework’s deployment-oriented Online Data portion.
Data
Data Lake: For use in constructing the prototype system, we obtained a variety of data
for a sample of customers making initial purchases between January 1, 2012, and March
1, 2014. As previously alluded to, incorporating rich relationship-oriented constructs
requires consideration of an array of structured and unstructured data sources: a non-
trivial task [13]. The various data sources incorporated in the system include structured
data from databases that support online transaction processing (OLTP) and CRM, as
well as unstructured data in the form of text log files, call center transcripts, and so on,
which are collectively referred to as a “data lake.”The Data Lake was generated by
obtaining these various raw data from CSI for a sample of 664,737 customers. As
depicted in Figure 4, the data lake included over 188 million raw data points.
Feature Generation: Once the data lake was constructed, the Feature Generation
component of the system was used to operationalize relationship-oriented constructs
identified in the framework. In order to simulate a realistic environment for imple-
mentation of the prototype system and avoid data leakage that could inflate accuracy,
we utilize a chronological rolling-window approach for creating training and test
sets. Only the first 30 days of customer history after initial purchase was used to
construct input features, and a period of 365 days to observe outcomes. In total, we
generated 1,003 features pertaining to the various construct categories. Transaction
and Demographics represent readily available baseline constructs. Transaction fea-
tures included order timestamps, amounts, prices, discounts, payment and shipping
methods, as well as purchased product information such as course names and topics.
556 KITCHENS, DOBOLYI, LI, AND ABBASI
Customers
Orders
Products
Transaction Support (OLTP)
Data Lake
Acquisition
Channels
Customer Interaction (CRM)
Voice of Customer
Feature
Generation
Transactions
Demographics
Channel
Product
Taxonomy
Construction
Messaging
Choice
Association
Rule Mining
Satisfaction
Engagement
Engagement
Measurement
Data
Preparation
Data Aggregation/
De-duplication
Missing Value
Removal
Outlier Removal
Discretization/
Binarization
Feature Type
Conversion
Feature Selection:
a. Remove Zero
Variance
b. Feature
Weighting
Data Matrix
Offline Modeling/
Evaluation
Machine Learning
Algorithm:
Composite
Convolution Kernel
SVM
Add-in/Leave-out
Analysis
Projected Cost-
Benefit Analysis
Windowing
(Temporal Effect
Adjustment)
Intelligent
Experimentation
Training/Test/
Validation Data
Balanced
Sampling
Top k% Churning
Evaluation
Online
Implementation
Modeling
Offline Evaluation
Field Experiments
A/B Testing
Online Evaluation
Short/Middle-Term
Cost-Benefit
Analysis
Analytics Infrastructure
Implementation
Big Data
Management
Iterative Process
Predictive
Model
Feature
Engineering
Pipeline
Big Data
Computing:
Streamlining/
Parallelization
Model Building
Platform
Relational
Database Files/Texts Intermediate
Data
IT Constructs Process Stage (dashed line means not
included in our study)
Satisfaction
Approximation
st/
ata
Campaigns
Marketing
Message
Log
Contact
Preference
Log
Message
Interaction
Log
Product
Reviews
Surveys/
Call
Center
Figure 3. System Diagram for Advanced Customer Analytics Instantiation
ADVANCED CUSTOMER ANALYTICS 557
Demographic features included age, gender, income, net worth, education, and
household information. With respect to relationship-oriented constructs, channel
variables were also relatively straightforward to operationalize. These features
describe the channels through which a customer was acquired and/or made an initial
purchase, including acquisition e-mails, call center upsells, paid search, partners,
prospect mailings, radio, social media, and television. However, other relationship-
oriented constructs such as choice, messaging, engagement, and satisfaction neces-
sitated the use of more involved logic and algorithms applied to multiple data
sources (as indicated in Figure 3). We discuss these construct categories in the
remainder of the section.
In a novel operationalization, our choice features focus on the interplay between
what products (and categories of products) a customer has purchased, relative to
what the company is offering and promoting. Figure 5 illustrates how the choice
variables were operationalized using a novel product taxonomy construction to
develop a tree of product categories, subcategories, products, and promotions related
to products purchased by the customer. Whereas prior studies have included only the
specific category of purchase as a variable, our taxonomic representation facilitates
richer contextualization of customer choice in the relationship—enabling enhanced
discriminatory potential. In the example, the purchased course “The Addictive
Brain”is used to generate features such as the number of choices in the same
category, subcategory, and promotions bundles received by the customer.
Engagement variables that are related to how a customer interacts with a company
provide essential intermediate cues regarding the status of the relationship [19]. However,
inclusioninpriorstudieshastypically been limited to service usage in industries such as
telecommunications. In the context of CSI, e-mail and clickstreams were prominent
avenues for customer engagement. Using the message interaction logs in our data lake,
we developed variables related to various engagement actions, including opening, for-
warding, and clicking e-mails, as well as viewing and engaging with the landing pages to
Figure 4. Summary Statistics of Data Related to Customer Churn Analytics
558 KITCHENS, DOBOLYI, LI, AND ABBASI
which they lead. Utilizing contact preference logs, we also developed engagement vari-
ables related to customer’s current contact preferences, as well as changes in preferences.
Satisfaction features focus primarily on online product reviews provided by
customers. CSI currently has no method for linking review authors to their customer
database, so in order to incorporate these data, a satisfaction approximation method
was employed. Review satisfactions (e.g., ratings, votes) and percentage change in
these measures were aggregated at the product level for products purchased by each
customer during the first 30 days of initial purchase. These contemporary review
characteristics for products purchased by the customer were used as proxy indicators
of satisfaction. As later demonstrated, this approach for overcoming the satisfaction
integration issue provided significant benefit in our system. We also included
measures of satisfaction and other perceptions from surveys and call center logs.
Afirm’smessaging to customers can have a profound impact on how customers
perceive the relationship [17,21] and on future customer behavior. Customers receive a
diverse set of mass and customized physical and digital messaging. Given that the mode,
quantity, and combination of messaging can impact customer behavior [59], we propose
a novel approach for extracting messaging patterns most likely to result in future
customer purchases. Adapting highly efficient new methods [65], we employed asso-
ciation rule mining to generate messaging features from hundreds of millions of messa-
ging records. We generated frequent item sets of messages, retaining only those sets that
culminated in a purchase in order to find messages strongly associated with purchases.
Details regarding this method are provided in online Appendix C. In addition to features
related to these association rule-mined messages, we included various other messaging
related features, including frequency of mail/e-mail, discounts, promoted products,
overlap with categories from a customer’s initial purchase, and so forth.
Data Preparation: Once the feature generation stage was completed, an initial data
matrix encompassing 1,003 variables for each of the 664,737 customers was con-
structed, comprising approximately 667 million values. Various necessary data
preparation steps were undertaken to handle veracity issues. Features or customer
Figure 5. Product Taxonomy and Choice Variables
ADVANCED CUSTOMER ANALYTICS 559
records with a high volume of missing values were removed. Outlier removal was
applied to eliminate records with abnormal entries (e.g., negative age or no transac-
tion history). Feature weighting by information gain and chi-squared statistics was
used to identify features with zero variance or negligible information. Ultimately,
435 features were retained. Moreover, random undersampling of the majority class
(i.e., churn) was used on the training data to achieve class balance.
Model
Offline Modeling/Evaluation: Kernel-based machine learning methods have garnered
attention from the information systems (IS) community in recent years for their ability
Figure 6. SVM Kernel-Based Customer Analytics Method
560 KITCHENS, DOBOLYI, LI, AND ABBASI
to derive patterns from large quantities of heterogenous, noisy, high-velocity data [3,4].
These methods have outperformed state-of-the-art rule-based, tree-based, Bayesian, and
deep learning models in recent benchmarking studies pertaining to voice-of-the-customer
tasks [4], while simultaneously providing the added benefit of greater transparency than
other big data machine learning methods through greater explanatory potential and
provisions for theory-driven design [2]. In sum, kernel-based methods afford the follow-
ing potential opportunities and benefits for our advanced customer analytics context:
●Principled, theory-driven kernel design by leveraging key customer nuances
elucidated by RMT
●Capability to effectively incorporate different types of customer–firm interaction
patterns manifested in diverse structured and unstructured enterprise data through
use of custom kernels designed for tabular, graphical, and string-based inputs
●Potential to efficiently fuse these diverse custom kernels through a meta-level
composite convolution kernel, providing robust and flexible ensemble-like
performance capabilities across a wide portfolio of customer analytics tasks
The novel kernel-based SVM customer analytics method we propose is depicted in
Figure 6. We begin with the RMT-based constructs previously described, including
engagement, satisfaction, choice, channel, and messaging. Next we leverage the afore-
mentioned enterprise customer analytics system encompassing an extensive offline data
environment coupled with an online real-time processing module (see Figure 3 for
details). In particular, the data lake, feature generation, and data preparation modules
of the system are used to convert raw customer relationship inputs into features and/or
formats suitable for our SVM composite convolution kernel. This results in three distinct
data representations. The tabular representation is composed of the 435 features detailed
in Figure 4. and Figure 4. The hierarchical representation utilizes the product portfolio
taxonomy (Figure 5), both in context of customer purchases and promotional offers. The
textual data incorporate four types of strings:
●Messages include all text of online/e-mail-based marketing initiated by the firm
●Engagement Logs include string representations of select customer interac-
tions with the firm, including mail/e-mail preference changes and open/click/
view activities
●Reviews include title and body text from reviews of products purchased by the
customer
●Product Info contains the title and description text of purchased products
The tabular, hierarchical, and textual representations feed into the composite convolu-
tion kernel, composing three underlying kernels: radial basis function (RBF), hybrid tree
(HT), and weighted cross entropy string (WCES). The main intuition guiding our
composite convolution setup is that customer–firm relationships embody dynamic,
multifaceted patterns that may encompass a plethora of manifestations, including
point-value quantifications of consumer decisions and actions, structural representations
of latent preferences, and semantic/stylistic indicators ofcustomer proclivities. A critical
ADVANCED CUSTOMER ANALYTICS 561
aspect of kernel-based methods is the kernel matrix comprising the similarity scores
between any two training instances.Next we describe eachof the underlying kernels at a
high level (online Appendix D includes additional details).
Using the tabular data representation, the RBF kernel uses a Gaussian classifier to
capture nonlinear “localized learning”patterns from point-value features [10]. The
HT kernel combines two tree methods [15]: a novel weighted shortest path tree
approach and a subtree method, both utilizing the hierarchical product line taxon-
omy. The shortest path approach utilizes two large global probabilistic trees encom-
passing all product purchases for positive and negative class cases observed in the
training set (e.g., churn and return). Link strengths between tree nodes are propor-
tional to product co-occurrence likelihood across all class instances in the training
set. Any two customers in the training set are compared on both trees by computing
the shortest paths between all nodes in their respective trees. To account for
potentially nuanced complementary/substitutive relations between product purchases
across customers, a purchase co-occurrence matrix is used to weight similarities. In
Figure 6, the Weighted Shortest Path illustration depicts an example involving two
customers (X and Y nodes), each with a two purchases (leaf nodes are products,
nonleaf nodes categories).
The second half of the HT kernel utilizes a labeled subtree method. Whereas the
shortest path approach is well-suited to capture probabilistic similarity between custo-
mer preferences, subtrees are effective in incorporating structural taxonomic similarity
[15] such as commonalities between customer preferences for certain specific cate-
gories, or category breadth versus depth similarities. Instead of relying on global trees,
the subtree method only compares the customer purchase trees. For instance, in the
Labeled SubTree example shown in Figure 6, X and Y gray customers each purchased
three items, including one common item (#3). The subtree method compares all unique
subtrees from each customers’purchase tree. Similarity between any two customers is
computed as the proportion of matching subtrees from their respective purchase trees. In
both the shortest path and subtree kernels, we also incorporate various constructs from
the tabular representation for each customer order item (i.e., a single block).
WCES is a novel string kernel [37] geared toward uncovering semantic and stylistic
customer tendencies hidden in text descriptions or logs. All training texts are tokenized
separately within each of the four categories of text, and unigram tokens are weighted
using the information gain heuristic. Due to class imbalance, undersampling, and use of a
single training set, we also incorporate a feature stability heuristic to as part of our token
weights to alleviate overfitting [33]. These weights are used as part of a cross entropy-
based customer similarity scoring mechanism. Cross entropy has been shown to be
effective in identifying commonalities in unstructured data distributions [30]. In our
context, it can help identify hidden commonalities in customer proclivities based on
purchase of introductory versus advanced products or certain styles or genres of offerings,
specific sticky promotional language, specialized purchase use cases, subtle communica-
tion and interaction indicators,andsoforth.AsdepictedinFigure 6, for each customer, we
randomly extract a predefined number of text windows of certain length (e.g., 50
562 KITCHENS, DOBOLYI, LI, AND ABBASI
characters) for each of the four textual representations. All such window pair combinations
between any two customers are compared using a weighted similarity comparison.
At the composite kernel stage, the RBF, HT, and WCES customer similarity kernel
matrices are fused using multiple kernel learning to allow robust predictive power
across an array of customer acquisition, retention, and expansion-related customer
analytics tasks. We employ this composite convolution kernel SVM model within
our prototype system to predict customer behaviors across a portfolio of applica-
tions. For comparison we also examined more traditional machine learning algo-
rithms, including stochastic gradient boosting, CART, boosted generalized linear
model, C5.0, random forest, and naive Bayes. Because of its ability to incorporate a
richer set of structured and unstructured data, the custom kernel SVM solution
described here significantly outperformed these methods. However, we found that
the more important factor for performance was the variety of data constructs
provided to the models. For further details, see online Appendix E.
Evaluation Metrics: To evaluate our system as well as determine the value of
individual data constructs, we created evaluation metrics based on the value of
actions taken as a result of model predictions. For the task of churn prediction, the
proposed action to be taken based on the model is to discontinue marketing physical
catalog mailings to the top 10 percent of customers most likely to churn. According
to CSI, the average cost of catalog mailings over the life of a customer is $66, and
the average lifetime revenue from a customer who does not churn after the first
purchase is $260. The costs and benefits of system-driven actions for the CLV and
conversion tasks are detailed in online Appendix B. Applied across an expected
population of 250,000 new customers each year, these figures are used to arrive at
the overall value for our system as well as each individual construct category.
Online Implementation: After building the prototype system, we delivered the system
and our results to CSI. They are currently testing to validate expected outcomes for
customers and creating a plan to implement the system in a production capacity. In
addition, based on the results of construct evaluation discussed next, CSI has already
made investments in infrastructure for data management and integration (discussed at
the end of the following section), which provides support for the core benefit of our
framework
Evaluation
Through evaluation of our prototype system implementation, we demonstrate the
utility as well as economic value of our contributed artifacts [49]. In the setting of
CSI, where customer outcomes are highly uncertain and costs of serving customers
are high, early predictions of customer behaviors are critical. Therefore, we use only
customer characteristics observable within the first 30 days of an initial purchase as
inputs to system predictions.
2
To ensure realistic results, we utilized chronological
evaluation with nine-month rolling windows for training, and one-month windows
for capturing test observations.
ADVANCED CUSTOMER ANALYTICS 563
Construct Evaluation: Customer Retention–Churn Application
Key features of our proposed framework for advanced customer analytics include: (1)
the comprehensive set of constructs to guide identification and acquisition of data for
advanced customer analytics applications; and (2) the mechanism for identifying the
value of these various constructs. By measuring the value of each individual construct,
IT managers asked to support big data analytics initiatives can make better-informed
decisions regarding investment in data management, integration, and real-time feature
engineering, and more effectively justify efforts in these endeavors. To accomplish
this, we trained models using (1) all construct categories,
3
(2) only transaction and
demographics as a baseline, (3) each individual category left out of the full model, and
(4) each individual category added in to the baseline. The results of these models are
presented in Tabl e 1. The full model provides 92.44 percent accuracy for the top 10
percent of customers most likely to churn, which would provide a net cost savings of
7.0 percent of CSI’s total marginal marketing spend. The baseline model, on the other
hand, has an accuracy of only 54.56 percent for the top 10 percent, resulting in a net
loss of revenue of 9.7 percent of total marginal marketing spend.
Table 1 reports z-scores for tests of differences in models’top 10 percent accura-
cies. Comparing add-in models to the baseline, each construct added significantly
Table 1. Add-In and Leave-Out Construct Evaluation
Accuracy
for top
10% most
likely to
churn
Annualized
net cost
savings, $
Net savings
as % of total
marginal
marketing
expense
Difference
from
baseline
model
(z-score)
Difference
from full
model
(z-score)
Baseline Model 54.56 –1,598,979 –9.69 –61.26***
Add-In Channel 57.22 –1,130,701 –6.85 3.82*** 57.92***
Add-In Satisfaction 67.84 –440,311 –2.67 19.45*** 44.01***
Add-In
Engagement
69.73 –317,802 –1.93 22.32*** 41.39***
Add-In Choice 71.44 –206,140 –1.25 24.96*** 38.96***
Add-In Messaging 78.46 250,079 1.52 36.15*** 28.29***
Leave-Out
Messaging
89.74 983,219 5.96 56.01*** 6.76***
Leave-Out
Satisfaction
90.02 1,001,085 6.07 56.54*** 6.12***
Leave-Out
Engagement
90.67 1,043,835 6.33 57.80*** 4.54***
Leave-Out Choice 91.74 1,113,385 6.75 59.88*** 1.84
Leave-Out
Channel
92.17 1,140,822 6.91 60.71*** 0.74
Full Model 92.44 1,158,688 7.02 61.26*** —
***p< 0.001; **p< 0.01; *p< 0.05.
564 KITCHENS, DOBOLYI, LI, AND ABBASI
improves the model accuracy. Engagement, choice, and messaging appear to be the
most valuable constructs from the add-in perspective, improving accuracy signifi-
cantly over the add-in satisfaction model, which, in turn, provides a significant
improvement over the add-in channel. The add-in comparisons are useful for
providing initial direction to determine which single construct could add the most
value if focused on first. From the leave-out perspective, engagement, satisfaction,
and messaging are the most important constructs. Depending on costs of data
management, integration, and real-time feature engineering of various constructs,
the model could be further tested with various construct subsets to determine the set
providing optimal value. It is important to note that all of the leave-out models
significantly outperform all of the add-in models, pointing to significant synergy
among the relationship-oriented constructs in making accurate predictions.
System Evaluation: Customer Retention–Churn Application
In addition to evaluating the set of constructs proposed and demonstrating the use of
our framework to search for a solution that provides optimal strategic value to the
organization, we compared our prototype system to several existing models to demon-
strate its value. As mentioned previously, no research has focused on models that
would be effective in high single-purchase, low frequency environments, which made
the search for benchmark models challenging. Almost all models created for churn
explanation and prediction rely on RFM constructs [20,52,55] that provide little
variation with which to predict in this environment, or perceptual constructs that are
difficult or impossible to measure for purposes of individual-level prediction [7,17].
Also as mentioned, much of the research examining churn prediction is focused solely
on innovation or choice in modeling techniques, with no regard to the constructs or
general framework for designing the analytics [36,60]. While this research is valuable,
it does not provide a valid benchmark for our prototype system, as our system is
composed of not only a novel method but also the set of relationship-oriented data
constructs supporting the method. For benchmark comparison, we had to find
approaches that encompassed a framework of constructs for prediction as well as a
model. Taking into account each of these challenges, we extensively surveyed the
literature to identify the most advanced, applicable approaches available.
The first benchmark identified was that of the classic RFM model. We implemented the
beta geometric/negative binomial distribution (BG/NBD) model of Fader et al. [20].
4
This
model is considered the standard for churn prediction, and it, or a less sophisticated
analogue, is widely used by many organizations for churn prediction (including CSI, prior
to this work). Next, we identified other models that also used RFM constructs, but added
additional features for prediction. Buckinx and Van den Poel [9] develop a churn predic-
tion system that relies on RFM constructs, other transaction characteristics, customer
choices, promotional messaging, and demographics and leverages logistic regression,
random forest, and neural network models. Coussement and De Bock [16] utilize general
additive models, decision trees, and random forests to predict churn using RFM and
ADVANCED CUSTOMER ANALYTICS 565
demographic characteristics. Neslin et al. [45] formulate an approach using logistic
regression with RFM and other transaction features along with messaging characteristics
to predict churn, specifically focusing on overcoming the “recency”trap, wherein many
customers have not purchased recently. Ballings and Van den Poel [6] use decision trees
and logistic regression to predict churn using RFM and other transaction features along
with demographics. We also identified models that used more perceptual constructs in the
tradition of Crosby and Stephens [17]. Specifically, Mittal and Kamakura [44]employ
measures of satisfaction combined with demographic characteristics to predict churn using
a probit choice model. Chen and Hitt [14] predict churn and switching behavior with a
logistic regression model including measures of satisfaction and engagement through
website usage characteristics as well as demographic information. Finally, Vanderveld and
Han [58] focus on engagement features to predict churn using random forests.
Each of these models employs a different prediction algorithm and set of constructs. In
order to compare them to our prototype system, we operationalized each construct
Table 2. Comparison to Benchmark Models
Model type
Accuracy
for top
10% most
likely to
churn
Annualized
net cost
savings, $
Net savings
as % of total
marginal
marketing
expense
Prototype system Composite
convolution
kernel SVM
92.4 1,158,688 7.0
Ballings and Van den
Poel [6]
Logistic regression 86.1*** 747,636 4.5
Decision tree 81.8*** 464,351 2.8
Bagged decision
trees
83.4*** 568,719 3.4
Buckinx and Van den
Poel [9]
Logistic regression 86.8*** 795,109 4.8
Random forest 83.4*** 570,460 3.5
Neural network 82.7*** 522,336 3.2
Chen and Hitt [14] Logistic regression 86.9*** 797,035 4.8
Coussement and De
Bock [16]
Generalized
additive models
86.2*** 755,287 4.6
Decision tree 83.1*** 549,650 3.3
Random forest 83.9*** 605,577 3.7
Fader et al. [20] BG/NBD 81.9*** 471,847 2.9
Mittal and Kamakura [44] Probit binary
choice
85.5*** 707,757 4.3
Neslin et al. [45] Logistic regression 84.4*** 634,536 3.8
Vanderveld and Han [58] Random forest 85.4*** 703,126 4.3
***Model performance inferior to prototype system at significance of p< .001.
566 KITCHENS, DOBOLYI, LI, AND ABBASI
included in their models based on data available for CSI. For instance, satisfaction
constructs were operationalized through online reviews contemporary to a customer’s
initial purchase, and engagement was operationalized through e-mail open and click
rates, as well as opt-in or opt-out of communications from CSI, as described previously.
Each model was evaluated using the same windowing strategy as the prototype system.
As shown in Table 2, the prototype system outperforms each of the benchmark
models by a wide margin. Based on z-tests for differences in model accuracies for
the top 10 percent most likely churners, the prototype system has significantly higher
accuracies than all other models at a significance of p< .001. The improvement of
the prototype over existing models is also highly economically significant, with net
savings of 7.0 percent of total marginal marketing spending, as compared to 2.9
percent to 4.8 percent saved by other models. This is despite the fact that the
benchmark models comprise leading approaches from marketing, IS, and machine
learning disciplines.
Construct and System Evaluation: Portfolio
Our prototype system includes a portfolio of three customer analytics applications:
●A churn prediction application focused on evaluating which customers the
firm should invest in through continued marketing efforts (retention)
●A conversion prediction application for identifying customers likely to
respond to individual e-mail promotions to reduce messaging fatigue and
prevent attrition (retention/expansion)
Table 3. Customer Analytics Portfolio Value, $
Value for Task
1: Churn
Value for
Task 2: CLV
Value for Task 3:
Conversion
Total
Value
Baseline Model N/A* 578,380 N/A* 578,380
Add-in Channel N/A* 715,453 N/A* 715,453
Add-in Engagement N/A* 929,421 N/A* 929,421
Add-in Satisfaction N/A* 1,053,120 N/A* 1,053,120
Add-in Choice N/A* 1,068,722 N/A* 1,068,722
Add-in Messaging 250,079 1,359,584 1,624,833 1,359,584
Leave-out Messaging 983,219 1,462,110 744,645 3,189,974
Leave-out Choice 1,113,385 1,739,599 1,717,484 4,570,468
Leave-out Satisfaction 1,001,085 1,562,407 2,180,741 4,744,233
Leave-out Engagement 1,043,835 1,584,695 2,366,044 4,994,575
Leave-out Channel 1,140,822 1,770,802 2,597,672 5,509,296
Full Model 1,158,688 1,784,175 2,759,812 5,702,675
*Model produces negative value; hence it would be foregone and therefore is not included in the
total value calculation.
ADVANCED CUSTOMER ANALYTICS 567
●A CLV prediction application for identifying customers who could success-
fully be expanded through offer of participation in a premium loyalty program
(expansion)
The CLV and conversion applications are detailed in online Appendix B. Table 3
demonstrates the total value achieved across the analytics portfolio as a whole, given
the inclusion or exclusion of various construct categories. Along with satisfaction and
engagement, which provided value to the churn prediction model, choice and messa-
ging are shown to have significant value across the portfolio of analytics applications.
This analysis illustrates how our proposed framework may be used to develop a
platform for agility in deploying customer analytics to achieve strategic value through
big data.
If, for example, investments in infrastructure were made to integrate all available
constructs into a feature generation pipeline for live deployment of the full model,
the estimated annualized value provided by the analytics portfolio would be
$5,702,675. This value (and that of future analytics) may be compared with the
costs of the infrastructure investment as a whole, as well as for each data construct/
source. For instance, if it is determined that the cost of obtaining and integrating data
for the engagement construct exceeds $708,100 (per annum), the firm may choose to
implement a system without it, resulting in a reduced $4,994,575 annualized value.
In addition to the various options shown in Table 3, other construct subsets may be
tested to choose the best complement of data. This evaluation provides value-based
justification for IT investment in data management and integration efforts to support
analytics initiatives.
Results Discussion
The instantiation of a portfolio of advanced customer analytics applications demon-
strates how our framework may be leveraged to develop capabilities for agility
through big data analytics and create strategic value and sustainable competitive
advantage in a dynamic market environment. By creating a value-justified infra-
structure for data integration and management to support advanced customer analy-
tics, firms can create a portfolio of analytics applications to serve a variety of
strategic purposes, providing significant value. The applications in our portfolio
combine to generate nearly $6 million annualized value and represent only a small
fraction of the opportunity for deploying analytics from this infrastructure to
strengthen and leverage customer relationships.
Evaluation of the system demonstrates that advanced customer analytics
systems built on relationship-oriented data from a variety of sources can
accurately predict customer behavior and add value. The prototype system
for churn prediction significantly outperformed each of the leading benchmarks
for comparison, including the well-adopted BG/NBD approach, evidencing the
impetus for relationship-oriented advanced customer analytics supported by IT-
enabled data infrastructure. Further, in addition to our system based on the
568 KITCHENS, DOBOLYI, LI, AND ABBASI
composite convolution kernel SVM, we also evaluated a system using more
traditional machine learning techniques (see online Appendix E), and found
that the SVM provides a significant lift in performance over the best alter-
native model, C5.0 (92.4 percent vs. 91.0 percent accuracy and $5.7 vs. $5.1
million net benefit), but variations in input data had a larger impact than
model choice.
Our results also demonstrate how, through use of the proposed framework,
the individual value of various data sources may be identified for use in
prioritization of data management and integration efforts. Each of the construct
categories identified aside from channel was shown to add significant value
over the portfolio of analytics applications. Engagement provides significant
value across all models, as suggested during construct justification because of
the information it provides regarding customers’continuing interest despite a
lack of transactions. When left out of the portfolio, it reduces value provided by
the analytics initiative by an estimated $708,100 per annum. Messaging and
choice, with rich information about customer preferences and communications,
each provide even more value ($2,512,701 and $1,132,207, respectively).
Satisfaction is also found to be important to the portfolio, which may be
expected due to its near ubiquitous prominence in the RMT literature [7,17,
44]. The $958,442 value is provided despite the fact that operationalizations
rely on contemporary review data to impute satisfaction levels for individual
customers rather than surveying every individual customer. There are also
significant synergies from the use of multiple data sources, which may be
seen by the value of the messaging construct. Alone it provides less than
$800,000 in value over the baseline, but its combination with other constructs
contributes over $2.5 million to the full model. This stresses the importance of
combining rich relationship-oriented data across silos, as well as the need for a
data valuation framework.
These values should be compared with data collection and integration costs in
order to guide investments in data infrastructure. The strength of results regarding
engagement and satisfaction constructs despite their limited available operationaliza-
tions have already led CSI to invest in changes to infrastructure to support data
management and integration. First, CSI has invested in a new platform for managing
its web presence in order to support the identification of individual customers
browsing its site and integration of these data into existing customer management
software. Prior to this work, CSI could analyze web traffic, but not tie it to individual
customers for performing analytics. Second, CSI has invested in infrastructure that
will allow it to tie online product reviews to the individual customers posting them.
Both of these changes were implemented as a direct consequence of the results of
this study, which provided IT management with the support and justification to make
these investments in data management and integration to support its advanced
customer analytics initiatives.
ADVANCED CUSTOMER ANALYTICS 569
Conclusion
In this study, we present a framework for designing advanced customer analytics
solutions based on relationship-oriented constructs. This work encompasses three
key contributions. First and foremost, we contribute to the design science literature
in the creation of a synergistic ecosystem of novel IT artifacts for performing
advanced customer analytics in the era of big data [28]. Our framework provides
guidance for the agile development and deployment of advanced customer analytics
solutions that predict customer behavior and inform strategic business decisions.
Following guiding principles from our framework, we develop a novel kernel-based
machine learning method that is custom-designed to extract insight and value from a
rich variety of relationship-oriented data constructs. We also provide a prototype
system instantiation of a portfolio of advanced customer analytics applications for a
firm with high proportions of single or infrequent purchase customers, a problem
that cannot be addressed by the siloed approaches of existing customer analytics
methods [6,41]. The system represents a rigorous proof-of-concept, while its results
offer practical relevance [28]. We show that this system enables significant strategic
value, contributing nearly $6 million in estimated annualized benefits. This value
will increase with additional analytics supported by the value-justified data infra-
structure informed by our framework.
Second, we contribute to managerial practice for firms attempting to employ big
data analytics to drive strategic value. Organizations are overwhelmed by available
data [40], and IT managers who are asked to support big data analytics through data
management and integration are in need of a blueprint for valuing various data
sources and justifying their efforts through return on investment in infrastructure [11,
35]. The framework provides a structure through which the various relationship-
oriented constructs can be evaluated based on added business value relative to costs
of data acquisition, management, integration, and real-time feature construction. Our
contributions in this area are validated by the direct impact of our results for CSI in
motivating strategic investment in data management and integration infrastructure.
Third, our research contributes to the nascent literature regarding predictive
analytics through the use of big data [2,5,13,22]. The kernel theory that we
employ in our design, RMT, has investigated many of the constructs that are central
to our framework [7,17,44]. However, all of this work has been from an explana-
tory, rather than predictive, viewpoint. We provide a firm foundation that allows us
to answer recent calls by many in the IS field for predictive analytics [56], particu-
larly utilizing volume and variety of available data [13,22] to predict micro-level
outcomes at the individual level [5,22]. The advanced customer analytics our
framework allows are not feasible in the absence of big data from a variety of
sources providing a relationship-oriented view of customer behavior.
In this era of profound digital transformation, customer agility lies at the intersection
of customer analytics, big data, and IT strategy. Firms capable of taking advantage of
such agility are best positioned to achieve sustainable competitive advantage. Our study
makes important contributions to the nascent literature on this critical topic.
570 KITCHENS, DOBOLYI, LI, AND ABBASI
Supplemental File
Supplemental data for this article can be accessed on the publisher’s website at
https://doi.org/10.1080/07421222.2018.1451957
NOTES
1. All studies cited in the references for the main body of our manuscript were reviewed, as
well as additional papers listed in online Appendix F (over 200 relevant studies in total).
While it is not feasible to review all relevant studies, this creates a representative set from
which we may draw conclusions about the completeness of our construct set.
2. We evaluated other observation period lengths (60 and 90 days), but the small improve-
ments in predictive power provided were outweighed by the reduction in value caused by
longer lead times to predictions.
3. No data for the firm and environment characteristics construct is included in our
prototype system. As noted in the description of this construct within the broader framework,
although potentially useful for prediction, this construct is likely more informative of shifts
suggesting models be revisited as firms monitor analytics solutions after deployment.
4. We would like to thank Peter Fader for his support as we implemented this model, as
well as for providing the data set used in his paper so that we could verify our implementation
to be identical.
REFERENCES
1. Abbasi, A.; Albrecht, C.; Vance, A.; and Hansen, J. Metafraud: A meta-learning frame-
work for detecting financial fraud. MIS Quarterly,36,4(2012), 1293–1327.
2. Abbasi, A.; Sarker, S.; and Chiang, R.H.L. Big data research in information systems:
Toward an inclusive research agenda. Journal of the Association of Information Systems,17,2
(2016), 1–32.
3. Abbasi, A.; Zahedi, F.; Zeng, D.; Chen, Y.; Chen, H.; and Nunamaker, J.F. Enhancing
predictive analytics for anti-phishing by exploiting website genre information. Journal of
Management Information Systems,31,4(2015), 109–157.
4. Abbasi, A.; Zhou, Y.; Deng, S.; and Zhang, P. Text analytics for sense-making in social
media: A language-action perspective. MIS Quarterly, forthcoming.
5. Agarwal, R., and Dhar, V. Big data, data science, and analytics: The opportunity and
challenge for IS research. Information Systems Research,25,3(2014), 443–448.
6. Ballings, M., and Van Den Poel, D. Customer event history for churn prediction: How
long is long enough? Expert Systems with Applications,39,18(2012), 13517–13522.
7. Bolton, R.N. A dynamic model of the duration of the customer’s relationship with a
continuous service provider: The role of satisfaction. Marketing Science,17,1(1998), 45–65.
8. Bolton, R.N.; Lemon, K.N.; and Verhoef, P.C. The theoretical underpinnings of custo-
mer asset management: A framework and propositions for future research. Journal of the
Academy of Marketing Science,32,3(2004), 271–292.
9. Buckinx, W., and Van den Poel, D. Customer base analysis: Partial defection of
behaviourally loyal clients in a non-contractual FMCG retail setting. European Journal of
Operational Research,164,1(2005), 252–268.
10. Burges, C.J. A tutorial on support vector machines for pattern recognition. Data Mining
and Knowledge Discovery,2,2(1998), 121–167.
11. Cappiello, C.; Francalanci, C.; and Pernici, B. Time-related factors of data quality in informa-
tion multichannel systems. Journal of Management Information Systems,20,3(2003), 71–91.
12. Chen, D.Q.; Preston, D.S.; and Swink, M. How the use of big data analytics affects
value creation in supply chain management. Journal of Management Information Systems,32,
4(2015), 4–39.
ADVANCED CUSTOMER ANALYTICS 571
13. Chen, H., and Storey, V.C. Business intelligence and analytics: From big data to big
impact. MIS Quarterly,36,4(2012), 1165–1188.
14. Chen, P.S., and Hitt, L.M. Measuring switching costs and the determinants of customer
retention in internet-enabled businesses: A study of the online brokerage industry. Information
Systems Research,13,3(2002), 255–74.
15. Collins, M., and Duffy, N. Convolution kernels for natural language. In Advances in
Neural Information Processing Systems,(2002), Vancouver, British Columbia, Canada. pp.
625–632.
16. Coussement, K., and De Bock, K.W. Customer churn prediction in the online gambling
industry: The beneficial effect of ensemble learning. Journal of Business Research,66,9
(2013), 1629–1636.
17. Crosby, L.A., and Stephens, N. Effects of relationship marketing on satisfaction, reten-
tion, and prices in the life insurance industry. Journal of Marketing Research,24,4(1987),
404–411.
18. Davenport, T.H. Analytics 3.0. Harvard Business Review, December 2013,64–72.
19. van Doorn, J.; Lemon, K.N.N.; Mittal, V.; et al. Customer engagement behavior:
Theoretical foundations and research directions. Journal of Service Research,13,3(2010),
253–266.
20. Fader, P.S.; Hardie, B.G.S.; and Lee, K.L. Counting your customers the easy way: An
alternative to the Pareto/NBD model. Marketing Science,24,2(2005), 275–284.
21. Gázquez-Abad, J.C.; Canniére, M.H. De; and Martínez-López, F.J. Dynamics of custo-
mer response to promotional and relational direct mailings from an apparel retailer: The
moderating role of relationship strength. Journal of Retailing,87,2(2011), 166–181.
22. Goes, P.B. Big data and IS research. MIS Quarterly,38,3(2014), iii–viii.
23. Goodhue, D.L.; Kirsch, L.J.; Quillard, J.A; and Wybo, M.D. Strategic data planning:
Lessons from the field. MIS Quarterly,16,1(1992), 11–34.
24. Goodhue, D.L., Wybo, M.D., and Kirsch, L.J. The impact of data integration on the
costs and benefits of information systems. MIS Quarterly,16,3(1992), 293–311.
25. Gunarathne, P.; Rui, H.; and Seidmann, A. Whose and what social media complaints
have happier resolutions? Evidence from Twitter. Journal of Management Information
Systems,34,2(2017), 314–340.
26. Gupta, S.; Hanssens, D.; Hardie, B.; et al. Modeling customer lifetime value. Journal of
Service Research,9,2(2006), 139–155.
27. Heudecker, N., and White, A. The data lake fallacy: All water and little substance.
Gartner, July 2014,6.
28. Hevner, A.R.; March, S.T.; Park, J.; and Ram, S. Design science in information systems
research. MIS Quarterly,28,1(2004), 75–105.
29. Hitt, L.M., and Frei, F.X. Do Better customers utilize electronic distribution channels?
The case of PC banking. Management Science,48,6(2002), 732–748.
30. Juola, P., and Baayen, H. A controlled-corpus experiment in authorship identification by
cross-entropy. Literary and Linguistic Computing, 20 (2005), 59–67.
31. Karimi, J., and Walter, Z. The role of dynamic capabilities in responding to digital
disruption: A factor-based study of the newspaper industry. Journal of Management
Information Systems,32,1(2015), 39–81.
32. Keane, T.J., and Wang, P. Applications for the lifetime value model in modern news-
paper publishing. Journal of Direct Marketing,9,2(1995), 59–66.
33. Koppel, M.; Akiva, N.; and Dagan, I. Feature instability as a criterion for selecting
potential style markers. JASIST,57,11(2006), 1519–1525.
34. Kunz, W.; Aksoy, L.; Bart, Y.; et al. Customer engagement in a big data world. Journal
of Services Marketing,31,2(2017), 161–171.
35. Laney, D. Why and how to measure the value of your information assets. Gartner,
August 2015.1–22
36. Lemmens, A., and Croux, C. Bagging and boosting classification trees to predict churn.
Journal of Marketing Research,43,2(2006), 276–286.
37. Lodhi, H.; Saunders, C.; Shawe-Taylor, J.; Cristianini, N.; and Watkins, C. Text
classification using string kernels. Journal of Machine Learning Research, 2 (2002), 419–444.
572 KITCHENS, DOBOLYI, LI, AND ABBASI
38. Lu, Y., and Ramamurthy, K. Understanding the link between information technology
capability and organizational agility: An empirical examination. MIS Quarterly,35,4(2011),
931–954.
39. Lyytinen, K., and Grover, V. Management misinformation systems: A time to revisit?
Journal of the Association for Information Systems,18,3(2017), 206–230.
40. McAfee, A., and Brynjofsson, E. Big data: The management revolution. Harvard
Business Review, October 2012,60–68.
41. Miglautsch, J. Application of RFM principles: What to do with 1–1–1 customers?
Journal of Database Marketing,9,4(2002), 319–324.
42. Miguéis, V.L.; Van den Poel, D.; Camanho, A.S.; and Falcão e Cunha, J. Modeling
partial customer churn: On the value of first product-category purchase sequences. Expert
Systems with Applications,39,12(2012), 11250–11256.
43. Mithas, S.; Ramasubbu, N.; and Sambamurthy, V. How information management
capability influences firm performance. MIS Quarterly,35,1(2011), 237–256.
44. Mittal, V., and Kamakura, W.A. Satisfaction, repurchase intent, and repurchase beha-
vior: Investigating the moderating effect of customer characteristics. Journal of Marketing
Research, 38 (February 2001)., 131–142.
45. Neslin, S.A.; Taylor, G.A.; Grantham, K.D.; and McNeil, K.R. Overcoming the
“recency trap”in customer relationship management. Journal of the Academy of Marketing
Science,41,3(2013), 320–337.
46. Neslin, S.A; Gupta, S.; Kamakura, W.; Lu, J.; and Mason, C.H. Defection detection:
Measuring and understanding the predictive accuracy of customer churn models. Journal of
Marketing Research, 43 (May 2006), 204–211.
47. Nunamaker, J.F.; Briggs, R.O.; Derrick, D.C.; and Schwabe, G. The last research mile:
Achieving both rigor and relevance in information systems research. Journal of Management
Information Systems,32,3(2015), 10–47.
48. Palmer, A. The evolution of an idea: An environmental explanation of relationship
marketing. Journal of Relationship Marketing,1,1(2002), 79–94.
49. Prat, N.; Comyn-Wattiau, I.; and Akoka, J. A taxonomy of evaluation methods for informa-
tion systems artifacts. Journal of Management Information Systems,32,3(2015), 229–267.
50. Ransbotham, B.S., and Kiron, D. Analytics as a source of business innovation. MIT
Sloan Management Review, February 2017,1–16.
51. Ransbotham, S.; Kiron, D.; and Prentice, P.K. Minding the analytics gap. MIT Sloan
Management Review, 56, (Spring 2015), 63–68.
52. Reinartz, W.J., and Kumar, V. The impact of customer relationship characteristics on
profitable lifetime duration. Journal of Marketing, 67 (January 2003), 77–99.
53. Roberts, N., and Grover, V. Leveraging information technology infrastructure to facil-
itate a firm’s customer agility and competitive activity: An empirical investigation. Journal of
Management Information Systems,28,4(2012), 231–270.
54. Sambamurthy, V.; Bharadwaj, A.; and Grover, V. Shaping agility through digital
options: Reconceptualizing the role of information technology in contemporary firms. MIS
Quarterly,27,2(2003), 237–263.
55. Schmittlein, D.C.; Morrison, D.G.; and Colombo, R. Counting your customers: Who are
they and what will they do next? Management Science,33,1(1987), 1–24.
56. Shmueli, G., and Koppius, O.R. Predictive analytics in information systems research.
MIS Quarterly,35,3(2011), 553–572.
57. Teo, T.S.H., and King, W.R. Integration between business planning and information
systems planning: An evolutionary-contingency perspective. Journal of Management
Information Systems,14,1(1997), 185–214.
58. Vanderveld, A., and Han, A. An engagement-based customer lifetime value system for
e-commerce. In 22nd ACM SIGKDD Conference on Knowledge Discovery and Data
Mining.2016.
59. Venkatesan, R., and Kumar, V. Framework for customer selection. Journal of
Marketing, 68 (October 2004), 106–125.
ADVANCED CUSTOMER ANALYTICS 573
60. Verbraken, T.; Verbeke, W.; and Baesens, B. A novel profit maximizing metric for
measuring classification performance of customer churn prediction models. IEEE
Transactions on Knowledge and Data Engineering,25,5(2013), 961–973.
61. Verhoef, P.C. Understanding the effect of customer relationship management efforts on
customer retention and customer share development. Journal of Marketing, 67 (October 2003).
62. Voss, G.B.; Godfrey, A.; and Seiders, K. How complementarity and substitution alter the
customer satisfaction–repurchase link. Journal of Marketing, 74 (November 2010), 111–127.
63. Wagner, C., and Majchrzak, A. Enabling customer-centricity using wikis and the wiki
way. Journal of Management Information Systems,23,3(2007), 17–43.
64. Walls, J.G.; Widmeyer, G.R.; and Sawy, O.A. El. Building an information system
design theory for vigilant EIS. Information Systems Research,3,1(1992), 36–59.
65. Wang, K., and Skadron, K. Association rule mining with the Micron Automata
Processor. IEEEInternational Parallel and Distributed Processing SymposiumHyderabad,
India. pp. 689–699, 2015.
66. Wixom, B., and Ross, J. How to monetize your data. MIT Sloan Management Review,
January 2017,10–13.
574 KITCHENS, DOBOLYI, LI, AND ABBASI
CopyrightofJournalofManagementInformationSystemsisthepropertyofTaylor&Francis
Ltdanditscontentmaynotbecopiedoremailedtomultiplesitesorpostedtoalistserv
withoutthecopyrightholder'sexpresswrittenpermission.However,usersmayprint,
download,oremailarticlesforindividualuse.