Journal of Intelligent Learning Systems and Applications, 2011, 3, 90-102
doi:10.4236/jilsa.2011.32011 Published Online May 2011 (http://www.SciRP.org/journal/jilsa)
Beyond Customer Churn: Generating Personalized
Actions to Retain Customers in a Retail Bank by a
Recommender System Approach
Michele Gorgoglione, Umberto Panniello
Department of Mechanical and Management Engineering, Polytechnic of Bari, Bari, Italy.
Received July 10th, 2010; revised October 1st, 2010; accepted December 1st, 2010.
ABSTRACT
Customer churn may be a critical issue for ban ks. The extant literature on statistical and machine lea rn ing fo r customer
churn focuses on the problem of correctly predicting that a customer is about to switch bank, while very rarely consid-
ers the problem of generating personalized actions to improve the customer retention rate. However, these decisions
are at least a s critical as the correct identification o f customers at risk. The decision o f what actions to d eliver to what
customers is normally left to managers who can only rely upon their knowledge. By looking at the scientific literature
on CRM and personalization, this research proposes a number of models which can be used to generate marketing ac-
tions, and shows how to integrate them into a model embracing both the analytical prediction of customer churn and
the genera tion o f reten tion actio ns. The b enefits and risks asso ciated with each a ppro ach a re discu ssed. Th e pap er also
describes a case of application of a predictive model of customer churn in a retail bank where the analysts have also
generated a set of personalized actions to retain customers by using one of the approaches presented in the paper,
namely by adapting a recommender system approach to the retention problem.
Keywords: Customer Churn, Customer Retention, Personalization, Predictive Models, Recommender Systems
1. Introduction
Retail banks often deal with customer churn. Among the
several issues addressed by Customer Relationship Ma-
nagement (CRM), identifying the customers who are
about to quit the relationship with a company is one of
the most important in the financial services industry.
When competition becomes tougher, when laws decrease
either the barriers to entry or the customer’s switching
costs, or when a company aims at strengthening its posi-
tion in a new market, the issue of retaining customers and
avoiding customer churn becomes even more crucial.
In the last decade several computer science scholars
have tackled the problem of building accurate models to
identify customers at risk in a bank by using statistical,
machine learning and data mining approaches [1]. How-
ever, much literature only focuses on the problem of
correctly predicting that a customer is about to switch.
Very rarely the problem of generating personalized ac-
tions to improve customer retention rates is considered.
The decision of what actions to deliver to what customers
is normally left to managers who can only rely upon their
knowledge. However, these decisions are at least as cri-
tical as the correct identification of customers at risk. A
customer churn prediction model should be integrated
with a model to decide what personalized marketing ac-
tion to deliver to customers; otherwise the benefits of
using accurate predictive models would be lost.
The issue addressed by this research is to identify a set
of approaches to generating actions to retain customers,
once customers at risk are identified and profiled. We
believe that this issue is important for both businesses
launching customer retention campaigns and researchers
aiming at building models of customer churn. From a
business perspective, the problem is to support managers
in making the right decision on how to define personal-
ized retention actions. In fact, this decision has to con-
sider both organizational aspects, such as control and
efficiency, and technological aspects, such as the accu-
racy of the algorithms. The challenge for managers is to
integrate the results of a predictive model of customer
churn with the decision of delivering certain marketing
actions. The problem is how to associate the right action

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 91
Retail Bank by a Recommender System Approach
with the right customer. From a scientific perspective, a
comprehensive model of how to generate actions might
be helpful to develop more effective data mining and
statistical models to support CRM processes.
By looking at both the scientific literature on CRM
and personalization, this paper proposes a number of
approaches which can be used to generate personalized
marketing actions, and discusses benefits and risks asso-
ciated with each model. Finally, a case of application is
described. A retail bank developed a predictive model of
customer churn. The model was integrated to a method to
generate personalized actions aimed at customer reten-
tion. The method is based to the approach followed by
Recommender Systems. The results of this model are
compared to those obtained by the same company in a
formerly launched retention campaign.
2. Prior Research
The problem of customer churn has been analyzed by
scholars in the fields of marketing and analytical CRM,
by analyzing several aspects. For instance, [2] analyzes
the possible payoffs of retention and acquisition strate-
gies depending on the market structure. An analysis of
the switching costs is provided by [3]. Customer churn
has also been studied by analyzing CRM-related prob-
lems such as the effectiveness of loyalty programs [4]
and customer satisfaction [5].
Analytical models to predict customer churn have been
developed in several areas and industries as well. Mar-
keting models have been deployed to predict customers
at risk [6,7] as well as data mining models [8], especially
in telecommunication-related industries [9-11]. Specific
prediction models have developed also in banking and
finance [12,13].
This body of literature have mainly focused on the
problem of correctly predicting customer churn. Very
rarely the problem of how to generate personalized ac-
tions, once customer churn has been modeled, is consid-
ered. For instance, [14] states that a good customer churn
model should also provide managers with a way to gen-
erate personalized actions, although the research focuses
only on predicting customer churn. The need to integrate
the analysis of customer behavior with the generation of
actions in a comprehensive model is maintained by [15],
while [16] suggests to include customers found at risk of
leaving the company in the organization of events in or-
der to strengthen their relationships with the company.
As the literature on customer churn does not directly
face the problem of how to generate actions, and in order
to achieve a more general understanding of the problem,
we have broadened the focus of the literature analysis to
the research on personalization-related areas. The goal is
to review the approaches to deliver actions to individual
customers. A review of studies on personalization can be
found in [17]. The authors consider three main areas,
namely computer science and information systems (CS/
IS), marketing, management science and economics. It is
possible to extract from that review the works directly
related to the problem of generating personalized actions.
In the CS/IS area the main streams dealing with ac-
tions are Recommender Systems and Web Contents Per-
sonalization. In both cases the problem of generating
personalized actions is solved by processing individual
customer profiles which describe personal preferences. In
a recommender system the profile describes a customer’s
preference by including either a set of ratings to certain
products or information on customers’ previous pur-
chases [18]. The action is the recommendation of a prod-
uct and it is generated by analyzing the similarities be-
tween the profiles of customers who bought or rated a
product and those of customers who did not. These sys-
tems are the most common approaches to personalization
in real industrial settings ever since Amazon adopted a
Recommender System [19]. In the case of Web Content
Personalization the user’s profile includes statistics on
the usage of the Web and the action is generated by asso-
ciating a content with similarities with the user’s profile
[20].
Scholars in marketing have dealt with the problem of
generating personalized actions by studying competitive
personalized promotions [21], customization [22], the
personalization of services [23], and recommender sys-
tems. In all cases, except that of recommender systems
which was reviewed above, the problem is solved by
using optimization models that allow the analyst to find
the best price and/or product based on individual cus-
tomers response models. The response models are built
by collecting data on customers’ previous behavior.
The research in management science [24] and eco-
nomics has mainly faced the problem of generating per-
sonalized actions by studying personalized pricing and
mass customization [25]. The first stream falls in the
marketing approaches reviewed above. The second str-
eam mainly analyzes the benefits and costs of producing
as many products as the number of different customers’
individual preferences a company can identify in the
market [26]. In these models the action is offering of a
vast variety of products. Actions are generated by pro-
cessing customers’ preferences in an aggregated way, not
individually, while the choice of the specific product is
left to customers.
A few research streams can be added to the review
taken from [17] which do not directly deal with person-
alization, namely Direct Marketing and Knowledge Dis-

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 92
Retail Bank by a Recommender System Approach
covery in Databases (KDD). An extensive literature ex-
ists in the area of direct marketing dealing with targeting
problems. The largest body of knowledge is related to the
process of segmentation, targeting and positioning. It
includes several methods to define and select a target
market and position an offer, i.e. define a specific mar-
keting mix (or any other marketing action) to that spe-
cific market. Several scholars in marketing have warned
about the fact that a segmentation should be done only
aiming at a clear marketing goal [27], i.e. segments should
be built in order to deliver targeted actions. The research
on direct marketing has produced models to define and
deliver actions to more granular targets, even single indi-
viduals. For instance, “RFM models” represent custom-
ers by their Recency (i.e., the time since they have made
the last purchase), Frequency (i.e., how often they pur-
chase), and Monetary value (i.e., the average money they
have spent) in order to compute the probability to which
a customer will make the next purchase by a certain pe-
riod of time. RFM models have quickly become common
in direct marketing and have been extended to give rise
to more sophisticated models such as Automatic Interac-
tion Detection Models and Regression Scoring models
[28]. In the industrial applications of these models, man-
agers often define a very small set of actions, even one
action (e.g. proposing a discounted subscription to a new
service), and run the model in order to select a subset of
individual customers who represent the optimal target.
Knowledge Discovery in Databases (KDD) is a pro-
cess to find valid, new, understandable and useful pat-
terns in data [29]. This process can be seen as a possible
way to target actions to customers. If the rules derivable
from KDD are “actionable”, i.e. present information on a
customer useful to decide “what to do” in order to mod-
ify that customer’s behavior, then a manager can define
marketing actions tailored to that customer [30,31]. The
main limitation of these approaches is that either they are
not efficient, due to the big amount of discovered rules
which need human supervision, or actions are not per-
sonal, because when rules are aggregated to improve effi-
ciency the target becomes aggregated as well. Starting
from KDD, some attempts were done to define a general
method to mining actionable patterns in databases and
thus generating personalized actions based on data de-
scribing customers’ reactions to the company’s previous
offers [32].
A last remark is useful to highlight that many compa-
nies, especially in banking and finance, make use of front
office personnel to generate actions tailored to the per-
sonal characteristics of customers. In this approaches,
managers have to talk to customers, understand their pre-
ferences and propose them specific offers. Actions can be
decided a-priori or a-posteriori. In the first case the front
office managers select the right customers to whom a
certain action can be offered, while in the second actions
are decided by the front office managers, who have to be
given the necessary degree of autonomy, based on their
understanding of customers’ needs.
3. Approaches to the Definition of
Personalized Actions
By reviewing the literature, five main categories can be
identified to classify the approaches to the generation of
personalized actions. Each category represents a set of
homogeneous approaches which can be used to decide
what action should be delivered to what customers in a
CRM program. The approaches are reported in Table 1
and discussed below.
Computational approaches include all those approaches
that build a complete model of customers’ behavior, ac-
tions, and customers’ reactions based on information
stored in a data set. These approaches can use both data
mining [32] and optimization models [21-23]. An exam-
ple of applications in banking and finance is a retail bank
which stores the data related to promotion of stocks, and
the reactions of customers who might have purchased
those stocks or not. The fundamental condition that en-
ables the adoption of these approaches is the complete-
ness of data. If the data sets do not include, for instance,
the reaction of each customer to a certain marketing
campaign, then neither a mining algorithm nor an opti-
mization model can be run to generate personalized ac-
tions. Computational approaches allow a company to
fully automate the generation of actions based on cus-
tomers’ profiles, as no human decision is needed. A li-
mitation is that only marketing actions already launched
before can be considered in such approach. The full cov-
erage of customers may be another problem because
some customers’ reactions may remain unknown (for in-
stance, when a customer does not respond to a survey).
Similarity-based approaches are those used by Re-
commender Systems [19] and Web content personaliza-
tion methods [20]. This kind of approach assumes that
actions are related to customer preferences, preferences
may be inferred by customer profiles, and that either
similar customers behave similarly or similar actions
cause similar reactions. A “similarity-based” approach
does not require to store as much information as a com-
putational approach. Recording customers’ preferences is
enough, because it is assumed that the unknown prefer-
ences of a customer can be derived by identifying the
similarity with other customers. However, the twofold
condition of applicability of such approaches is that cus-
tomers’ profiles have to represent preferences, and only

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 93
Retail Bank by a Recommender System Approach
actions associated with those preferences can be gener-
ated. For instance, a customer who owns multiple credit
cards can be classified as a customer who “prefers” using
credit cards, whereas the fact that a customer has a mort-
gage does not necessarily represent a “preference”. A “si-
milarity-based” approach is useful to automate the per-
sonalization process. An example of this approach in
banking is given in Section 5.
Bottom-up approaches include the knowledge discov-
ery methods [31,33] and the use of front office personnel.
These approaches consist of two separate steps: 1) pro-
filing customers, 2) deciding proper actions, where the
first step has to precede the second step (i.e., actions de-
pend on profiles). They cannot be fully made automatic
because only the first step is performed by an algorithm.
For this reason these approaches are typically not very
efficient. The condition of applicability is that the deci-
sion-making effort has not to exceed the company’s re-
sources: either the number of customers is low or the
number of decision-makers is high. The advantage is that
targeting can be very effective because each profile is
thoroughly analyzed before generating a proper action.
An example of bottom-up approach in banking is the
work of a financial advisor who manages a portfolio of
customers. The advisor has periodic conversations with a
customer, analyzes her needs and proposes tailored fi-
nancial solutions.
Top-down approaches include the direct marketing
approaches [28]. They consist of the same two separate
steps typical in bottom-up approaches. However, in this
case, the decision of what actions to deliver is made be-
fore the definition of customers’ profiles and, hence, pro-
files depend on actions. For instance, a retail bank man-
agers may first decide to offer customers a discount on
bank transfers, and then select the target customers by
building appropriate profiles. For this reason top-down
approaches do not cover the whole customer base: only
customers falling in the target profiles are delivered the
actions. In this approach, the control on the decision-
making process is very high, as the definition of actions
is made centrally (it is not delegated to front office per-
sonnel as in a bottom-up approach). Profiles can be de-
termined by either running algorithms or asking the front
office personnel to identify suitable target customers.
The condition of applicability is that actions can be de-
fined before customers’ profiles. In some CRM problems,
such as reducing customer churn, this condition can be
questionable.
Customization approaches are those which offer cus-
tomers many different options and let the customers
choose the suitable one. This approach is typically adopt-
ed in mass customization [25]. In this approach actions
do not derive from processing customers’ profiles but
rather from customers’ choices. This makes customiza-
tion different from other personalization approaches [26].
The offer has to be granular enough in order to adopt this
approach. For instance, a retail bank can propose cus-
tomers to choose one among many discounted options
(e.g., a discount on bank transfers, credit cards, cash
cards, e-transactions, etc.). The association between ac-
tions and profiles is performed by the customers instead
of the company. For this reason, the control over the
process is low and the resulting actions can turn out to be
quite expensive. However, the targeting is expected to be
quite effective.
4. Which Approach is Best? Discussion of
Benefits and Risks
The approaches identified above have characteristics that
make them beneficial in certain settings while risky in
others. The suitability of each approach should be thor-
oughly discussed by managers based on the observation
of their benefits and risks compared to the company
goals and business conditions. Typically, managers have
to deal with trade-offs. The main characteristics which
can differentiate the five approaches are the following:
 control on the decision-making process;
 automation of the personalization process;
 effectiveness of targeting;
 process efficiency;
 scope of actions;
 coverage of customers.
4.1. Control on the Decision-Making Process
Companies normally want to keep as much control as
possible on the definitions of the actions to deliver. Pos-
sible reasons are the need of controlling the costs related
to marketing actions, and avoiding to leave too much
power to front office personnel. In this case, the decision
of what offering to customers should be centralized. As
an example, a retail bank may want to avoid that the front
office personnel makes the decision of what to offer to
the customers in their portfolio, because both costs may
increase and the bank could be exposed to opportunistic
behavior. In this case, the suitable approaches to be used
are “computational”, “similarity-based” and “top-down”.
In all these approaches, the decision of what action to
deliver is made centrally, by the support of either an al-
gorithm or the strategic orientation. The “bottom-up” and
“customization” approaches may be risky because the
first may involve the use of front office personnel, the
second requires the intervention of customers.
4.2. Automation of the Personalization Process
The need of automating the process of profiling, defining

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 94
Retail Bank by a Recommender System Approach
personalized offers and delivering actions can come from
several reasons, such as the need of keeping the cost low
and the promptness of CRM processes high. For instance,
companies in the financial services industry may want to
be prompt and fast in starting a retention program when
the churn rate suddenly increases. In this case, automat-
ing the process is critical. “Computational” and “similar-
ity-based” approaches are the most suitable when the
need of automation is high. In fact, they can be imple-
mented into an algorithm which can be integrated in the
company information system. The remaining approaches
(“bottom-up”, “top-down” and “customization”) can only
be partially automated, as it was observed in the previous
section.
4.3. Effectiveness of Targeting
Companies involved in CRM programs are often con-
cerned in maximizing targeting, i.e. to maximize the
probability that a customer receives an offer tailored to
her own needs. For instance, a retail bank may want to
make sure that each customer is offered exactly the pro-
duct she needs. In these case, the best approaches to de-
fine personalized actions are the “computational”, “bot-
tom-up” and “customization” ones. In the first approach
actions are derived by knowing the customers’ response.
In the second, actions are defined by analyzing each pro-
file. In the third, customers themselves decide what ac-
tion they prefer. Targeting is less effective in the “simi-
larity-based” and “top-down” approaches because the
customers’ need are inferred by similarity or simply not
considered when the decision of actions is made. The
effectiveness of targeting can be evaluated by the “re-
demption rate” which measures how many customers
react the expected way to a marketing campaign.
4.4. Process Efficiency
Whatever the goal of a CRM program, companies have
to keep the costs of CRM as low as possible. Companies
may choose the best approach to define personalized
actions by considering the level of resources they can
allocate to that process. Some approaches requires less
resources than others. Particularly, “similarity-based”
and “top-down” approaches look suitable to this aim. The
algorithms required by the first approach are relatively
simple, as it is witnessed by their quick diffusion in in-
dustrial settings. The second approach makes the integra-
tion between the decision of actions and definition of
profiles simpler than other. On the contrary, the algo-
rithms required by a “computational” approach are more
complex. A “bottom-up” approach requires either a huge
effort in terms of supervision of the knowledge discovery
process or a high level of resources in terms of people
involved in the process. The efficiency of a “customiza-
tion” approach seems to depend on the business settings,
as it can be easy to implement but the results can turn out
to be expensive to the company. Efficiency can be meas-
ured by the number of people involved and/or by time
spent on the process.
4.5. Scope of Actions
Companies may want to be creative in the definition of
the actions to deliver to customers. The proposed approa-
ches have different impacts on the capability of the com-
pany to define a wide spectrum of actions. The only ap-
proaches that look suitable to this aim are those in the
“bottom-up” category. In fact, it always requires the hu-
man supervision, to either infer the needs from the pro-
files and, in turn, the actions, or talk to each customer,
dig into his/her preferences and needs, and finally find
out the best offer to deliver. The scope of actions in
“computational” and “similarity-based” approaches are
limited by the nature of the data. In the former, the only
actions which can be offered to customers are those re-
corded in the databases, for which customers’ reactions
are already recorded and codified. In the latter, actions
have to be related to the variables used to profile cus-
tomers. In the “top-down” approach, the scope of the
actions is usually small because companies define a
small set of offers. In the “customization” approach the
scope of actions depends on the spectrum of different
choices customers are provided with. A potential meas-
ure for this variable can be the number of different ac-
tions.
4.6. Coverage of Customers
Not all approaches allow a company to fully cover the
customer base with personalized actions. The “similar-
ity-based” and “bottom-up” approaches guarantee a full
coverage. In the first, actions can be defined by similarity.
Even customers for whom the amount of data is very
little (e.g., new customers), actions can be derived by the
preferences of similar customers. In the second, human
decision makers can rely on all their knowledge to define
actions starting by profiles. On the contrary, in a “com-
putational” approach the coverage is limited to customers
for whom relevant information on actions and reactions
have been recorded. In a “top-down” approach, actions
are defined independently of profiles, with the conse-
quence that not all profiles can be associated with actions.
In a “customization” approach only customers who make
an explicit choice are then offered the actions, those who
do not want to express any preference remain outside the
coverage of the CRM program. The coverage of custom-
ers can be measured by the percentage of customers for

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a
Retail Bank by a Recommender System Approach
95
whom actions can be derived from profiles.
Table 1 represents each approach in terms of its char-
acteristic, conditions that must be true to adopt that ap-
proach, the benefits and risk that should be discussed in
order to select the most appropriate approach.
5. A Case of Application
Born as a small local bank, the company went through a
significant M&A process over the last few years. It ac-
quired many new customers as a result of this process,
currently having about 300 000 customers in total and
being a medium-sized Italian retail bank operating at a
national level. In 2008 the bank managers had observed a
growing churn rate. Although the churn rate was much
smaller than the acquisition rate, the cost for acquiring a
new customer was much higher than the cost of retaining
a customer. Moreover, the cost for the re-acquisition of a
lost customer was even higher. These observations made
the need of reducing customer churn one of the bank’s
strategic priorities.
At the end of 2008 a team of managers was involved
in a project aimed at both identifying customers at risk of
leaving the bank and launching a marketing campaign to
retain those customers. By talking to financial advisors
and senior managers, the team identified a number of
events which normally signal that a customer is about to
leave the company. The most important signals were a
relevant decrease in the assets, the sale of more than one
financial products formerly owned, the cancellation of all
the automatic outgoing payments and incoming credits.
By analyzing the customer data set, the managers identi-
fied the customers at risk by selecting those customers
who presented one or more signals of imminent attrition.
They were ordered by the number of signals of attrition
in a ranking from high to medium risk.
Once the customers were identified, the team had to
carefully face the problem of what actions should be part
of a marketing campaign. A fairly straightforward choice
would be leaving the decision to the personal financial
advisors, who are sales reps with a portfolio of about 150
(on average) customers to manage. Since they know each
customer personally, they could decided the best offer to
retain each customer. However, the bank top manage-
ment decided that the actions to deliver to customer had
to be identified centrally. The reason was threefold. Fir-
stly, not all customers had a direct and strong relation-
ship with their personal advisors. Some preferred to in-
teract with the bank via the home banking system or
other virtual channels. Secondly, the high number of
personal advisors, about 2000, would make the process
quite expensive. The auditing process would be slow and
complex. Thirdly, the top managers wanted to keep a
strong control over the whole customer relationship
management process. The problem was too crucial to be
left to the initiative of front office personnel.
Therefore, in order to generate appropriate retention
actions, the managers had several meetings with senior
managers and sales/marketing managers. A list of few
marketing actions was generated, such as offering the
customer a credit card for free or a discount on the an-
nual fee. The list was communicated to all the personal
financial advisors. They were told to pick one or more
actions among those in the list for each customer identi-
fied at risk of leaving the bank, depending on the cus-
tomer’s profile. The actions had to be suitable to each
specific customer and picked based on the advisors’
Table 1. Approaches to the definition of personalized retention actions.
Approach Characteristics Conditions Benefits Risks
Computational Complete model of preferences,
actions and reactions
The data set includes information on
prior marketing actions and
customers’ responses
• Control
• Automation
• Targeting
• Limited scope of actions
• Customer coverage
Similarity-based Comparison between customer
profiles
Actions are related to customer
preferences, and preferences may be
inferred by customer profiles
• Control
• Automation
• Efficiency
• Coverage
• Limited scope of actions
Bottom-up
The definition of customers’
profiles precedes the definition
of actions
The supervision effort has to be
affordable (in terms of number of
profiles and resources)
• Targeting
• Scope of actions
• Coverage
• Low efficiency
• Lack of control
Top-down
The definition of customers’
profiles follows the definition
of actions
Actions can be defined independently
of customers’ profiles
• Control
• Efficiency
• Lack of targeting
• Limited coverage of
customers
Customization Customers are free to choose
the appropriate action The offer has to be granular enough • Targeting • Low efficiency • Lack of control

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 96
Retail Bank by a Recommender System Approach
knowledge of the customer’s behavior.
As a result, a redemption rate of 78.5% was observed,
meaning that 1062 customers out of 1353 contacted cus-
tomers responded positively to the marketing action
proposed and did not leave the bank, thus reducing the
overall churn rate.
Although the results were not impressive, they were
fairly aligned with the managers’ expectations. In fact,
the manager themselves were aware of some flaws in the
first project. First of all, actions were independent of
customers’ profiles. The decision of what to offer to cus-
tomers at risk was made upstream, based on previous
experiences rather than a deep knowledge of each custo-
mer. Therefore, a very small number of customers would
receive an offer tailored to their actual needs. Another
flaw of the method was the very limited possibility to
automate the process. While the identification of custo-
mers at risk could be implemented on an information
system, the decision-making process through which to
identify retention actions could not.
Encouraged by the results and the commitment of the
bank executives, the team decided to launch a new, im-
proved project of customer retention. The main goal
would be improving targeting and the possibility of au-
tomating the process.
5.1. Predictive Model of Customer Churn
The managers decided to collaborate with a group of
analysts, including this paper’s authors. The project fol-
lowed the Cross Industry Standard Process for Data Mi-
ning (CrispDM) methodology [34]. The first step was
aimed to codify the managerial knowledge and beliefs, to
define organizational and business constraints and to
identify the sources of useful information. This was help-
ful to guide the analysts in the definition of the variables
that would be included in the model. For instance, in the
managers’ experience the more the products owned by a
customer the higher the probability that she would be
loyal, while customers owning only one product had the
highest probability of leaving the company. The products
which made customers loyal were typically financial
products such as loans, stocks and bonds.
In the next step, the analysts discussed with the man-
agers the nature of the data sets to use based on the first
step output. The data warehouse service was outsourced
by the bank. Therefore, any data set had to be explicitly
asked to the outsourcer which managed the data storage.
Firstly, the team decided to focus only on customers in
the “basic” segment with at least a charged bank account.
The “basic” segment was the one with the highest num-
ber of customers and the highest churn rate. Customers in
higher segments shall not be included in the analysis.
Secondly, the group decided not to include the data re-
lated to customers in the branch offices most recently
acquired from other banks, in order to keep the data ho-
mogeneous. Thirdly, the data would be referred to the
last two years. The reason was that the bank had older
data only for some customers. The data available on cus-
tomers coming from other banks recently acquired by
M&A processes was not older than two years. Moreover,
a technical constraint in the storage service made the
retrieval of data referred to years preceding the last two
very slow. Finally, the data would include demographical
information on customers, such as age and job, the his-
tory of the relationship between customers and bank (e.g.,
when an account was open or when a financial product
was purchased and/or sold), the products owned by cus-
tomers (e.g., credit cards, cash cards, stocks, insurances,
etc.), the operations made on the account. Operations
were classified and divided into subcategories: 1) pay-
ments by card, bank transfers and cheques (incoming and
outgoing), 2) payments to utility services, 3) salaries and
similar incomes credited on the account, 4) other opera-
tions. Other information for each customer were also
included in the data set, such as the name of the cus-
tomer’s personal financial advisor, the branch office the
customer belongs.
Once the group was provided with a data set, a pre-
liminary analysis was run in order to check the correct-
ness of the data. The acquisition and analysis proceeded
in an iterative way, until a complete and correct set of
data was provided. However, the data on operations on
customers’ accounts was available only for the last year.
For this reason that specific data was not used in the
subsequent modeling phase. The initial dataset included
around 200 000 customer, 300 000 records and 100 va-
riables.
The next step was aimed to clean and modify the data
according to the analysis’ goals and method. Some re-
cords were deleted while others were re-coded, and seve-
ral variables were transformed starting from the initial
data set. The records referred to users older than 70 years
were dropped from the database (customers older than 70
could leave the bank for “natural” reasons more than for
a real choice). The records referred to customers with
more than one account were merged into one overall re-
cord, in order to obtain a data set where each customer
corresponded to one record. Several variables were
transformed. Some were discretized (e.g., length of the
relationship between customer and bank). For some va-
riables the analysts computed statistics such as average,
maximum and minimum value (e.g., operations made on
the account). Mistakes were cleaned (e.g., double codes
for the same value of a variable) and some records with

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 97
Retail Bank by a Recommender System Approach
too many missing values were deleted.
The final number of customers in the data set was
203 196. More than 200 variables were included in the
new data set, although only a small subset would be used
in the subsequent modeling phase. The depth of historical
data owned by the bank was two years, namely 2007 and
2008. During the analysis which was performed in the
first months of 2009, the bank continued to gather similar
data in order to apply the model to a new set of data.
The core step of the project was aimed to build a pre-
dictive model of customer churn and test it on actual data.
After several experiments, a relatively small subset of 27
variables, taken from the initial set of 200, was selected
to build the model. The set included the length of the
relationship between customer and bank, the customer’s
place of residence, the type of card owned (e.g., cash,
credit or pre-paid card), the type of financial products
and services owned (e.g., stocks, loans, insurances) and
their quantity and value, the type of account owned (cor-
responding to a certain set of discounted options on the
account), the amount of payments and incomes auto-
matically debited and credited, respectively, on the ac-
count (see Table 2).
After several trials, the algorithm used to build the
predictive model was J4.8. The training and validation
sets included all data referred to 2007 and 2008. Each
point was represented by the figure at the end of each
semester in 2007 and 2008, therefore 4 vectors of 27 va-
riables for each customer were used at most to train the
model. A ten-fold cross-validation method was used.
This data included information about 11,000 customers
who had abandoned the bank between 2007 and 2008
and about 180,000 customers who had remained loyal in
the same period. An additional test was performed by
applying the model, once trained and validated, to the
data related to the first semester of 2009. This data set
was gathered while the analysts were building the model
and the test consisted of applying the selected model,
with the selected settings, to the actual data in order to
predict the churner users.
Two predictive models were built by the analysts and
provided to the bank managers who would select one,
based on some managerial implications. The models pre-
dicted whether a customer would leave the company or
not, based on the data. In fact, because of the limited
temporal depth, it was not possible to train the model to
predict customer churn within a specific period of time.
The output of the models were two lists of customers
with a predicted binary value (“will leave” or “will be
loyal”) and a corresponding score. The first model se-
lected 11,070 potential churners among the active cus-
tomers with an accuracy equal to 87.35% and a true posi-
Table 2. Variables used in the modeling phase.
Variable Type
Length of customer-bank relationship numerical1
Place of residence nominal
Type of card owned nominal2
Maximum value of a customer’s transaction numerical
Type of transaction corresponding to the maximum
value nominal
Maximum quantity of a product purchased numerical
Type of product corresponding to the maximum
quantity purchased nominal
Type of account owned by the customer nominal3
People sharing the same account with the customer binary
Products owned by the customer nominal4
Automatic payment of utility services binary
Automatic credit of incomes on the account binary
1All the numerical variables were discretized; 2split in 4 binary variables;
3split in 3 binary variables; 4split in 11 binary variables.
tive rate equal to 0.916. The second model selected 32 938
potential churners among the active customers with an
accuracy equal to 67.34% and a true positive rate equal
to 0.994. The potential “churners” identified by the first
model were a subset of those selected by the second
model, except 282 customers (0.14% of the whole cus-
tomer base).
The bank’s managers decided to select the second
model. The decision was based on the fact that the cost
of a false negative was considered substantially higher
than the cost of a false positive. The managers preferred
to contact more customers, even loyal customers mis-
classified as potential “churners”, than risking not to
identify actual customers at risk. Moreover, the second
model performed significantly better than the first, once
tested on the first semester of 2009, as the second identi-
fied 96.89% of customer who actually left the bank,
whereas the percentage of re-identification in the first
model was 79.33%.
5.2. Selection of a Suitable Approach to the
Generation of Personalized Actions
The next problem was to integrate churn prediction with
the decision of which marketing actions should be deliv-
ered to which customer. The analysts proposed all the
five approaches presented in the section above to gener-
ate personalized actions for those customers predicted as
potential “churners”. Firstly, the team discussed the con-

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a
Retail Bank by a Recommender System Approach
98
ditions of applicability of each approach. Secondly, the
main managerial issues were ranked by priority in order
to discuss benefits and risks.
paigns had been carried out by defining non-personalized
actions, and the goal was just to improve targeting. A
“top-down” approach would represent too small a step
towards personalization, while a “similarity-based” ap-
proach would allow the bank to achieve a better targeting.
Figure 1 reports a graphical representation of the process
followed to select the most suitable approach to the defi-
nition of personalized actions.
The only condition to discard, among those listed in
Table 1, was the one related to the “computational” ap-
proach. Prior actions and customers’ response data had
never been stored in the dataset by the bank. Given the
absence of this kind of data, a “computational” approach
could not be adopted. All the other approaches could be
potentially adopted. Actions could be inferred by cus-
tomer profiles by a “similarity-based” approach, at least
partially, because they included some behavioral data
(products owned). The bank could afford the supervision
of effort in a “bottom-up” approach if the process was
delegated to front office personnel. Actions could be de-
fined independently of customers’ profiles, in a “top-
down” approach, because managers had some experience
on prior campaigns. Finally, the offer was granular
enough to let the customers choose in a relatively wide
set of alternatives, thus adopting a “customization” ap-
proach.
5.3. Generating Retention Actions by a
Recommender System
The most common practical applications of the “similar-
ity-based” approach are Recommender Systems (RS).
Given a set of customers and a set of products or items, a
RS predicts the unknown utility of an item for a customer.
If item j is predicted to be of a high utility for customer i,
then the system delivers the marketing action of recom-
mending that customer to purchase that item. The action
is personalized because the set of items recommended to
customers is different for each customer.
More formally, a RS deals with two types of entities:
users (e.g., customers) and items (e.g., products). In the
transaction-based RSs, the utility of an item for a custo-
mer is measured by a Boolean variable indicating if the
user owns a particular item or not, or with the purchasing
frequency of an item, or with the usage frequency of that
item [35,36]. Based on the known values of utility, a RS
tries to estimate the utility of the yet unseen items for
each customer. In other words, a RS can be viewed as the
rating function R that maps each user/item pair to a par-
ticular utility value [18]:
The discussion of benefits and risks was crucial to se-
lect the most appropriate approach. The two main issues
raised by the managers were ranked as (1) keeping high
control on the marketing process, (2) achieving better
targeting with respect to prior marketing campaigns.
According to the first issue, the team excluded the adop-
tion of both a “bottom-up” and a “customization” ap-
proaches. In fact, given the high number of customers
predicted at risk, a “bottom-up” approach would be fea-
sible only by delegating the task of revising each profile
and deciding the corresponding actions to the front office
personnel. This would entail a substantial loss in control.
The same would happen by letting customers choose
among different optional actions in a “customization”
approach. The choice between a “similarity-based” and a
“top-down” approach was eventually made based on the
second issue. In fact, the bank’s prior marketing cam-
R: Users
× Items  Utilities (1)
One of the main tasks of a RS is to make this function
total by estimating the unknown utilities. The estimation is
done by using a similarity function, which can be designed
by following one of the following three approaches [18]:
No data on
actions and
reactions
What
conditions are
not true?
Exclude the
“computational”
approach
Control
Exclude the
“bottom-up” and
“customization”
approaches
Targeting
Exclude the
“top-down”
approach
Select the
“similarity-based”
approach
What are the
main issues?
Figure 1. Representation of the decision-making process.

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 99
Retail Bank by a Recommender System Approach
 Content-based recommendations: the user is rec-
ommended items similar to the ones the user pre-
ferred in the past;
 Collaborative-filtering recommendations: the user
is recommended items that people with similar
tastes and preferences liked in the past;
 Hybrid approaches: they combine collaborative-
filtering and content-based methods.
This general approach had to be adapted to the prob-
lem of generating retention actions to customers at risk of
churning. Three hypotheses were identified in order to
consider the approach feasible: 1) recommending certain
products to customers can improve retention; 2) the pro-
ducts able to improve customer retention are not the
same for all customers and can be identified by looking
at the behavior of loyal customers; 3) similar customers
behave similarly and have similar preferences.
While the third hypothesis has much support in the
marketing literature, the first two were significantly
supported by the management expertise and partially
by empirical evidences. In fact, the bank’s managers
were confident that certain products, such as bank’s
stocks, insurance products and loans, directly contrib-
ute to improve customers’ loyalty. These beliefs had
emerged in the first project step in the statements of
several managers interviewed by the analysts. The
managers were also persuaded that the products which
can improve retention are not the same for all custom-
ers, as each customer may have a different sensitivity
to different products. These differences come from
several causes, such has different personality traits and
different life cycle stages. Because of this variety of
behavior, the statistical analyses only partially vali-
dated these beliefs. Customers had been split into loyal
and those who had abandoned the bank, although some
differences could be observed between the two groups
in the distribution of the products owned by customers,
they were statistically significant only at a low prob-
ability.
As a consequence of these observations, the team of
analysts and managers decided to adopt a collaborative
filtering approach to recommendations, where each
customer at risk would be recommended products that
loyal customers with similar tastes and preferences
liked in the past.
Each user was defined as a vector containing both de-
mographic and transactional data. That kind of informa-
tion was used because two users can be defined as simi-
lar only if they are similar in terms of preferences (e.g.,
products and services owned by customer, quantity and
amount of transactions) and other behavioral and demo-
graphical variables (e.g., length of relationship with the
bank, place of residence).
According to the standard collaborative filtering ap-
proach, the “neighborhood”, i.e. the set of customers z
with tastes and preferences similar to customer i, was
formed by using the cosine similarity, which is given by
the following formula:
 
,,
22
22 ,,
,cos, iz
iz iz
is zs
sS
is zs
sS sS
rr
simi zrr



 



iz
iz iz (2)
where ri,s and rz,s are the ratings of item s assigned by
user i and user z respectively. Siz={s  Items|ri,s ≠  
rz,s ≠ } is the set of all items co-rated by both user i and
user z, and

iz denotes the dot-product between the
vectors i and z. In this specific case, the customer i be-
longs to the set of customers predicted as churner,
whereas the customer z belongs to the set of loyal cus-
tomers Z. Only the customers predicted as loyal with a
probability higher than 90% were considered in Z. The
strong hypothesis of this approach is that similar cus-
tomers behave similarly, so a customer predicted to leave
the company can be retained if she is recommended
products characterizing the loyal customers with similar
tastes and preferences.
Therefore the main difference between a standard re-
commender system approach and the system used in this
case is the way the neighborhood is formed. While in the
former, the neighborhood include all the customers simi-
lar to the customer to whom an action has to be delivered,
in the latter the customers included in the neighborhood
were taken from a subset of all customers, namely the
most loyal customers. The final output of the recom-
mendation algorithm was a list of products that should be
suggested to each customer at risk.
As a result, a redemption rate of 81.6% was observed,
meaning that 1313 customers out of 1609 contacted cus-
tomers, responded positively to the marketing action
proposed and did not leave the bank, thus reducing the
overall churn rate.
5.4. Discussion of Results
The method used in the first project can be classified as a
hybrid of a “top-down” and a “bottom-up” approaches.
In fact, on the one hand the team of managers decided
what actions should be delivered to customers independ-
ently of the profiles of customers at risk. The decision
was made by talking to senior managers and was based
on a general knowledge of the way customer reacted to
marketing campaigns in the past. On the other hand, the
team asked the front office personnel (the personal fi-
nancial advisors) to select the most appropriate action

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 100
Retail Bank by a Recommender System Approach
among those identified by senior managers. This selec-
tion would be based on each customer’s profile.
The flaws of this method mainly came from those as-
sociated with a “top-down” approach. In fact, the front
office personnel (the “bottom-up” component of the pro-
cess) was involved only downstream and had a small role.
In fact, the main flaws were the lack of targeting and the
impossibility of automating the process. Despite the
flaws, the method had some benefits, again those associ-
ated with a “top-down” approach. Firstly, the bank had
the control over the process, because the decisions were
made centrally. Secondly, the process was relatively effi-
cient compared to other methods, because the main re-
source spent was the time of junior and senior managers.
The second project should then improve targeting and
automation without decreasing control and efficiency.
The method used in the second project can be classi-
fied as a “similarity-based” approach. The possibility of
defining actions directly by computing similarity was
given by the fact that some of the variables used to define
customers’ profile were “actionable”. Particularly, the
ownership of products (cards and financial products, see
Table 2) could be used to derive actions, i.e. the offer of
products not owned by the target customer but owned by
loyal customers similar to the target customer.
The results reached by the second project were consis-
tent to the expectations, at least partially. Control and
efficiency did not decrease. In fact, the decisions were
made again centrally by the team. The involvement of
the front office personnel was no longer necessary, not
even to select the most appropriate action, because ac-
tions were already tailored to each customer’s profile.
The time spent to talk to senior managers was approxi-
mately the same. However, a fair comparison of the effi-
ciency is not possible, because in the second project the
team used the knowledge gained in the first to select the
appropriate variables to build the customers’ profiles and
draw actions. Therefore, the time spent with senior man-
agers was very small, because they only had to approve
the set of actions defined by the team.
Automation remarkably improved. At the end of the
project, the variables to build the customers’ profiles
were defined, the data mining model was set, and the
algorithm to compute the actions by similarity was codi-
fied. At this stage, the team had only to implement a pro-
cedure to get the right information from the databases
and run the algorithms periodically.
The level of targeting only slightly increased. Trying
to explain this result is helpful to highlight the limits of a
“similarity-based” approach. Two main reasons can ex-
plain the observation.
First of all, a customer’s decision of either leaving a
company or remaining loyal is affected by many factors.
Some are related to the specific products the company
offers to customers. For instance, a customer may find
more convenient the offer of a competitor and decides to
switch. Some are related to the relationship between cus-
tomer and company. For instance, a customer may feel
like she is not treated the right way or the relationship is
not as intense as it should be in order to curb the decision
of leaving. Other factors pertain to customers’ personal
conditions and situations, such as when a customer de-
cides to quit because of a weak financial situation sug-
gesting to save money. Given this behavioral complexity,
a “similarity-based” approach makes the very simple hy-
pothesis that the decision of remain loyal only depends
on improving the company offerings.
The second reason is that a “similarity-based” app-
roach is intrinsically characterized by a low “scope of
actions”, as it was defined in Section 4, and this ends up
to affect the level of targeting. The number of different
actions that such approach can identify is limited by the
nature of the data. In the case described above, the ac-
tions were associated with the variables describing whe-
ther the customer owns a product (cards or financial pro-
ducts). Therefore, the company can only offer a customer
one or more products she does not own yet, chosen
among those owned by similar customers predicted as
loyal. This generated a very limited set of feasible ac-
tions, far away below the number of possible combina-
tions of the set of products. Because of this intrinsic li-
mitation, the number of different actions the company
could offer was so lower than the number of customers,
that many customers received the same offer given to
other customers. Actually, the set of actions coming from
the application of the “similarity-based” method was
very similar to the actions the managers had proposed in
the first marketing campaign. The consequence of this
limited targeting can explain the fact that several cus-
tomers did not react positively to the marketing action, so
reducing the overall redemption rate. The slight increase
in the redemption rate was probably due to a better pre-
dictive model of customer churn rather than to a better
method used to associate the actions with the profiles. In
fact, while the first method was based on a set of “alert
signals”, the second had learnt a customer behavior mod-
el. However, the results obtained by the “similarity-based
method” still remain interesting, because they demon-
strate the applicability of the method in a real industrial
setting.
6. Conclusions
The issue tackled by this paper is how to integrate the
problem of predicting customer churn through analytical

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 101
Retail Bank by a Recommender System Approach
models with the problem of generating personalized ac-
tions to retain customers. The two problems have been
treated separately by research so far, as the extant litera-
ture witnesses. On the one hand, research has developed
accurate statistical and machine learning models to asso-
ciate each customer’s profile with a churn score. On the
other, the decision of what personalized action to deliver
to each customer has remained a problem for managers.
This paper presented a list of possible approaches to ge-
nerate personalized actions and discussed the relationship
between the generation of actions and the customer pro-
filing. Benefits and risks associated to each approach are
also discussed. A case of application was described to
support the theoretical discussion. In a retail bank, a team
of analysts developed a predictive model of customer
churn and integrated it with a method to generate person-
alized actions to retain customers. Actions were gener-
ated by using a “similarity-based” approach. The specific
approach was selected by firstly discussing the condi-
tions of applicability of each approach and secondly the
main business priorities. This allowed the analysts to
discard the unfeasible approaches, and select the most
appropriate one. In the business case described, the per-
sonalized actions for customer retention were generated
by adapting the algorithm of a recommender system. The
main difference between a standard recommendation en-
gine and the one used in the application case is the way
the users neighborhood was formed, namely by comput-
ing the similarity between customers predicted at risk of
churn and loyal customers.
The results were aligned with the expectations, at least
partially. In fact, the method was expected to improve
targeting and automation, without decreasing control and
efficiency. Control and efficiency remained the same,
while automation remarkably increased. The level of tar-
geting did not increase as much as expected, because of
two main reasons. The first is that a complex behavior,
the customer’s decision of leaving or remaining loyal, is
reduced to a much simpler hypothesis, namely the fact
that the decision only depends on what products the
company offers. The second is that the scope of actions
in a “similarity-based” approach is quite limited, and this
intrinsically reduces the level of targeting. Further re-
search will be done to reduce these limits and to extend
the comparison to other approaches outlined in the paper.
REFERENCES
[1] J. Hadden, A. Tiwari, R. Roy and D. Ruta, “Computer
Assisted Customer Churn Management: State-of-the-Art
and Future Trends,” Computers and Operations Research,
Vol. 34, No. 10, 2005, pp. 2902-2917.
doi:10.1016/j.cor.2005.11.007
[2] N. B. Syam and J. D. Hess, “Acquisition Versus Reten-
tion: Competitive Customer Relationship Management,”
Working Paper, University of Houston, Houston, 2006.
[3] P.-Y. Chen and L. M. Hitt, “Measuring Switching Costs
and the Determinants of Customer Retention in Inter-
net-Enabled Business: A Study of the Online Brokerage
Industry,” Information Systems Research, Vol. 13, No. 3,
2002, pp. 255-274. doi:10.1287/isre.13.3.255.78
[4] R. N. Bolton, P. K. Kannan and M. D. Bramlett, “Impli-
cations of Loyalty Program Membership and Service Ex-
periences for Customer Retention and Value,” Journal of
the Academy of Marketing Science, Vol. 28, No. 1, 2000,
pp. 95-108. doi:10.1177/0092070300281009
[5] T. Hennig-Thurau and A. Klee, “The Impact of Customer
Satisfaction and Relationship Quality on Customer Reten-
tion: A Critical Reassessment and Model Development,”
Psychology and Marketing, Vol. 14, No. 8, 1998, pp.
737-764.
[6] S. A. Neslin, S. Gupta, W. Kamakura, J. Lu and C. H.
Mason, “Defection Detection: Measuring and Under-
standing the Predictive Accuracy of Customer Churn
Models,” Journal of Marketing Research, Vol. 43, No. 2,
2006, pp. 204-211. doi:10.1509/jmkr.43.2.204
[7] Z. Jamal and R. E. Bucklin, “Improving the Diagnosis
and Prediction of Customer Churn: A Heterogeneous
Modeling Approach,” Journal of Interactive Marketing,
Vol. 20, No. 3-4, 2006, pp. 16-29. doi:10.1002/dir.20064
[8] J. Hadden, A. Tiwari, R. Roy and D. Ruta, “Churn Pre-
diction: Does Technology Matter?” International Journal
of Intelligent Technology, Vol. 1, No. 1, 2006, pp. 104-
110.
[9] S. V. Nath and S. Behara, “Customer Churn Analysis in
the Wireless Industry: A Data Mining Approach,” Pro-
ceedings of the 34th Meeting of the Decision Sciences In-
stitute, Washington, 22-25 November 2003, pp. 505-510.
[10] L. Bin, S. Peiji and L. Juan, “Customer Churn Prediction
Based on the Decision Tree in Personal Handyphone
System Service,” Proceedings of International Confer-
ence on Service Systems and Service Management,
Chengdu, 9-11 June 2007, pp. 1-5.
[11] J. Lu and O. Park, “Modeling Customer Lifetime Value
Using Survival Analysis—An Application in the Tele-
communications Industry,” Data Mining Techniques,
SAS Users Group International, Vol. 28, 2003, pp. 120-
128.
[12] T. Mutanen, “Customer Churn Analysis: A Case Study,”
Research Report, No. VTT-R-01184-06, 2008.
[13] D. V. D. Poel and B. Larivière, “Customer Attrition
Analysis for Financial Services Using Proportional Haz-
ard Models,” European journal of Operational Research,
Vol. 157, No. 1, 2004, pp. 196-217.
doi:10.1016/S0377-2217(03)00069-9
[14] S. Y. Hung, D. C. Yen and H. Y. Wang, “Applying Data
Mining to Telecom Churn Management,” Expert Systems
with Applications, Vol. 31, No. 3, 2006, pp. 515-524.

Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a
Retail Bank by a Recommender System Approach
102
doi:10.1016/j.eswa.2005.09.080
[15] L. S. Yang and C. Chiu, “Knowledge Discovery on Cus-
tomer Churn Prediction,” Proceedings of the 10th World
Scientific and Engineering Academy and Society: Inter-
national Conference on Applied Mathematics Dallas,
Texas, 1-3 November 2006, pp. 523-528.
[16] C. B. Bhattacharya, “When Customers are Members:
Customer Retention in Paid Membership Contexts,”
Journal of Marketing Science, Vol. 26, No. 1, 1998, pp.
31-44.
[17] G. Adomavicius, Z. Huang and A. Tuzhilin, “Personal-
ization and Recommender Systems,” Tutorials in Opera-
tions Research, INFORMS, Charlotte, 2008.
[18] A. Tuzhilin and G. Adomavicius, “Towards the Next
Generation of Recommender Systems: A Survey of the
State-of-the-Art and Possible Extensions,” IEEE Trans-
actions on Knowledge and Data Engineering, Vol. 17, No.
6, 2005, pp. 734-749.
[19] G. Linden, B. Smith and J. York, “Amazon.com Recom-
mendations: Item-to-Item Collaborative Filtering,” IEEE
Internet Computing, Vol. 7, No. 1, 2003, pp. 76-80.
doi:10.1109/MIC.2003.1167344
[20] B. Mobasher, R. Cooley and J. Srivastava, “Automatic
Personalization Based on Web Usage Mining,” Commu-
nications of the Association for Computing Machinery,
Vol. 43, No. 8, 2000, pp. 142-151.
doi:10.1145/345124.345169
[21] J. Zhang and M. Wedel, “The Effectiveness of Custom-
ized Promotions in Online and Offline Stores,” Journal of
Marke ting Research, Vol. 46, No. 2, 2009, pp. 190-206.
doi:10.1509/jmkr.46.2.190
[22] A. Ansari and C. F. Mela, “E-Customization,” Journal of
marketing research, Vol. 40, No. 2, 2003, pp. 131-145.
doi:10.1509/jmkr.40.2.131.19224
[23] J. Pancras and K. Sudhir, “Optimal Marketing Strategies
for a Customer Data Intermediary,” Journal of Marketing
Research, Vol. 44, No, 4, 2007, pp. 560-578.
doi:10.1509/jmkr.44.4.560
[24] B. P. S. Murthi and S. Sarkar, “The Role of the Manage-
ment Sciences in Research on Personalization,” Man-
agement Science, Vol. 49, No. 10, 2003, pp. 1344-1362.
doi:10.1287/mnsc.49.10.1344.17313
[25] A. Kumar, “From Mass Customization to Mass Personal-
ization: A Strategic Transformation,” International Jour-
nal of Flexible Manufacturing Systems, Vol. 19, No. 4,
2007, pp. 533-547. doi:10.1007/s10696-008-9048-6
[26] N. Arora, X. Dreze, A. Ghose, J. Hess, R. Iyengar, B.
Jing, Y. Joshi, V. Kumar, N. Lurie, S. Neslin, S. Sajeesh,
M. Su, N. Syam, J. Thomas and Z. Zhang, “Putting
One-to-One Marketing to Work: Personalization, Cus-
tomization and Choice,” Marketing Letters, Vol. 19, No.
3, 2008, pp. 305-321. doi:10.1007/s11002-008-9056-z
[27] J. H. Myers, “Segmentation and Positioning for Strategic
Marketing Decisions,” South-Western Educational Pub,
Cincinnati, 1996.
[28] W. J. Reinartz and R. Venkatesan, “Decision Models for
Customer Relationship Management,” In: B. Wierenga,
Ed., Handbook of Marketing Decision Models, Springer
Netherlands, Dordrecht, 2009.
[29] U. Fayyad, G. Piatetsky-Shapiro, P. Smith and R. Uthu-
rusamy, “Advances in Knowledge Discovery and Data
Mining,” The Massachusetts Institute of Technology
Press, Massachusetts, 1996.
[30] A. Silberschatz and A. Tuzhilin, “On Subjective Meas-
ures of Interestingness in Knowledge Discovery,” Pro-
ceedings of the 1st International Conference on Knowl-
edge Discovery and Data, Montreal, 20-21 August 1995,
pp. 275-281.
[31] B. Mobasher, B. Berendt and M. Spiliopoulou, “Knowl-
edge Discovery and Data for Personalization,” Tutorial at
the 12th European Conference on Machine Learning,
Freiburg, 5-7 September 2001.
[32] K. Wang and Y. Jiang, “Mining Actionable Patterns by
Role Models,” IEEE International Conference on Data
Engineering, Atlanta, 3-7 April 2006, pp. 16-25.
[33] A. Tuzhilin and G. Adomavicius, “Export-Driven Valida-
tion of Rule-Based User Models in Personalization Ap-
plications,” Data Mining and Knowledge Discovery, Vol.
5, 2001, pp. 33-58. doi:10.1023/A:1009839827683
[34] C. Shearer, “The CRISP-DM Model: The New Blueprint
for Data Mining,” Journal of Data Warehousing, Vol. 5,
No. 4, 2000, pp. 13-22.
[35] Z. Huang, W. Chung and H. Chen, “A Graph Model for
E-Commerce Recommender Systems,” Journal of the
American Society for Information Science and Technol-
ogy ,Vol. 55, No. 3, 2004, pp. 259-274.
doi:10.1002/asi.10372
[36] Z. Huang, X. Li and H. Chen, “Link Prediction Approach
to Collaborative Filtering,” Proceedings of the 5th ACM/
IEEE-CS Joint Conference on Digital Libraries, Denver,
7-11 June 2005, pp. 14-142.