flexible payment recommender system - rutgers …cimic.rutgers.edu/~soon/papers/jita2011.pdf ·...
TRANSCRIPT
Journal of Information Technology and Architecture
Vol. 8. No. 4, December 2011, Pages 299-316
299
Flexible Payment Recommender System
Soon Ae Chun1*, Yoo Jung An2, Sunju Park3, June-Suh Cho4 and James Geller5
1City University of New York – College of Staten Island2Fairleigh Dickinson University
3Yonsei University, Seoul, Korea4HanKook University of Foreign Studies, Seoul, Korea
5New Jersey Institute of Technology
(Received November 10, 2011; Revised November 25, 2011; Accepted December 3, 2011)
Abstract: When a customer owns multiple credit cards, the selection of a credit card for a purchase
payment is rarely done systematically. Furthermore, the customer has no online options to pay a high
cost item with multiple cards. This paper presents a flexible Payment Recommender (PR) system that
helps the customer to make a decision on identifying a group of credit cards, called Virtual Card, which
provides customers with the flexibility of customizing their payments by splitting a payment over
multiple cards. This intelligent PR system uses a divisible card payment decision model based on fuzzy
membership functions of the customer’s weighted preference parameters and suggests the best card or
best combination of cards to the customer. The PR system uses the Virtual Card Manager (VCM) that
modifies and extends the existing single credit card-based payment infrastructure, to support processing
multiple card-based flexible payments. A prototype PR system was implemented, that can be installed
on the customer’s PC or mobile device, to manage multiple credit cards and personal preferences, and
to support making payment decisions. A user acceptability study of the proposed PR system is also
reported.
Keywords: payment recommender system, divisible card payment decision model, flexible card,
virtual card, virtual card processing system, e-commerce, m-commerce
1. Introduction
A typical online payment involving a credit-card
payment consists of the following steps: the user
selects a credit card and charges the whole amount
for his/her purchase on the chosen card; the card
information is sent over a secure connection such as
SSL, and the credit is verified with the card provider.
In the case of a mobile payment with a credit card,
a mobile device acts as a credit card where a credit
card can be entered either directly to the device (e.g.
enter the credit number on the cell phone), or
through a card swipe or wireless card scanning with
the mobile device. Once the credit card information
is entered, the financial gateway interacts with the
card provider and verifies credit card authentication
and approval. The majority of the current electronic
commerce transactions or mobile payments are still
based on one credit card for one purchase.
However, a customer typically owns a set of credit
cards to choose from for payment. Each credit card
has certain features associated with it, such as inter-
est rates, credit limit, as well as some promotional
features such as cash-back on a percentage of total
purchases made, travel protection insurance, extend-
ed warranty on purchases, frequent flier miles, etc.
The customer is faced with making the decision on
which credit card to use for a particular transaction.
Typically, the customer’s choice of one credit card
over another (e.g. an Amex card over a Master card)
is arbitrary; for instance, she chooses one card over
another based on the first one at hand or based on
a habit. At best, her decision is made based on a
vague sense of the credit card provider’s reputation,
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
300 Journal of Information Technology and Architecture
remaining balance or interest rate, if she can keep
track of the features and remember them. In the cur-
rent payment scenarios, there is no automated sup-
port for a flexible payment system that recommends
the best card for each transaction that would opti-
mize her preferences and spending habits, consider-
ing all the combinations of criteria, benefits and
features provided by different cards.
Furthermore, the current credit card based online
payment methods lack flexibility that includes the
ability to split a purchase price over several cards,
instead of charging it on a single card. This flexibility
of using multiple cards for one payment is especially
beneficial for the customer. For example, a customer
may have a preferred card A with a low interest rate
and a second card B with a high interest rate. With-
out an option of splitting the purchase price over sev-
eral cards, the credit limit and prior spending on A
may force her to pay the purchase with the higher
interest rate card, B. The customer, however, would
be better off if she could first max out the remaining
credit limit on card A and then pay the balance of the
purchase price with card B, lowering her high inter-
est balance on card B. In other cases, a customer may
face a problem when paying for the purchase of a
major item, if its price is larger than the available
credit on each of her credit cards. While the total
available credit on all her cards may well exceed the
sale price, the merchant would not be able to com-
plete the sale, because today’s e-payment systems
can only accept a single credit card for a single trans-
action. In this case, it is in the best interest of the
merchant to allow the customer to split the payment
over multiple credit cards.
How to split a purchase price over multiple cards,
however, is a complex decision making problem [4].
Today's typical customer carries a wide variety of
cards with a corresponding variety of incentives,
such as frequent flyer miles, cash back, introductory
low interest rates, promotional gifts, etc. In addition,
some credit cards are issued through employer orga-
nizations, called secondary card issuers, with com-
pany specific policies, such as that the card is allowed
only for business expenses or for office supplies,
which will further affect the choice of cards. Fur-
thermore, sellers may impose their own constraints
such that not every credit card is accepted every-
where. Lastly, the choice of the ideal card (or com-
bination of cards) often depends on the paying habits
of the customer and the calendar date of the pur-
chase. Most customers would prefer to pay with a
credit card, which will buy them a delay. Thus, all
other factors being equal, a customer would prefer to
use a card where the closing date is 20 days away
over a card where the closing date is only 2 days
away. For customers who carry large balances over
long periods, the interest rate is an important deci-
sion factor. For the customers who pay their bills off
every month and do not care about interest rates at
all, on the other hand, the closing date or other incen-
tive offers might be more important parameters.
Therefore, while it is clearly in the interest of a
credit card user to be able to split a purchase price
over multiple cards, as described above, she may find
herself in a state of confusion about how to split the
payment exactly. We call this problem of choosing an
optimal card or card combination as the divisible
payment decision problem. Since the criteria for
choosing a set of credit cards for payment are diverse
and differ from one customer to another, customers
need help in keeping track of all their cards and their
incentives and constraints. In addition, at the time
of making a purchase, customers need help with the
complex decision of which card(s) to use and how to
optimally split a purchase over multiple cards if one
chooses to use multiple cards.
In this paper, we provide a flexible payment rec-
ommender system that solves the divisible payment
decision problem and proposes to the user the flex-
ible payment choice with the best combination of
cards for one purchase, based on the cards’ properties
such as promotional incentives, interest rates, as
well as users’ preferences. In order to accommodate
these flexible payment methods with multiple cards,
we propose a divisible card payment infrastructure
that modifies and extends the current credit card
payment verification infrastructure, often used in
electronic commerce. First, to support the divisible
card payments, the Virtual Card Manager (VCM) is
added to the merchant side. The VCM handles the
Flexible Payment Recommender System
Journal of Information Technology and Architecture 301
divisible card approval process between the mer-
chant and the respective credit card issuers. Second,
to support the customer’s card-usage decisions, the
new infrastructure provides the customer with a
Payment Recommender agent (PR) that supports the
customer’s payment decision. The PR is added to the
customer PC or a mobile device. By modeling the cus-
tomer’s preferences using weighted fuzzy set mem-
berships, the fuzzy payment recommender agent
considers the preferences over the card issuers’ pol-
icies, such as credit limits, interest rates and many
others, as well as the policies imposed by the sec-
ondary issuers, such as employers, to suggest the
best combination of cards to the customer.
We present a prototype implementation of a pay-
ment recommender system, which allows split credit
card payments and helps the customer optimize the
distribution of his charges over multiple cards. The
payment recommender system can also be installed
on a client’s mobile device which can recommend
payment over one or more credit cards, which can be
particularly suitable for m-payments. For example, a
customer shopping in a store should able to use a
PDA to compute an optimal card split and then
transfer the solution by infrared or Bluetooth signal
to the cash register and finalize the transaction.
Contributions of this paper include (1) a flexible
payment recommender model to support the credit
card choices for a purchase that is based on fuzzy
membership of various card characteristics and pol-
icies, and user preferences; (2) It supports payment
with multiple credit cards for a purchase, which is
novel in the literature. Typically the current status
is to support micropayments, but our model could
support purchases with large payment amounts with
multiple cards, and support users with limited
credit; (3) The payment infrastructure with multiple
card transactions has been designed with minimal
changes to the existing payment infrastructure; (4)
A prototype system has been implemented to illus-
trate the feasibility of the proposed model; and (5) A
user survey was conducted to evaluate the user
acceptability of the flexible recommender system.
The rest of the paper consists of the following sec-
tions. The background review of the related litera-
ture is provided in Section 2. Section 3 describes our
divisible payment decision model based on fuzzy set
theory. Section 4 proposes the new divisible card pay-
ment infrastructure while comparing it with the
existing infrastructure, followed by the prototype of
the Payment Recommender system in Section 5. In
Section 6, we present and discuss the results of a
user evaluation of the flexible payment concept, and
Section 7 concludes the paper.
2. Related Work
Mobile payments (m-Payments) are defined as
payments that a mobile device carries out for e-com-
merce and/or standard commerce. M-commerce allows
such monetary transactions over a wireless network
and is estimated to reach $88 billion of the total glo-
bal m-commerce revenue by 2009 [12]. The number
of m-commerce transactions is expected to increase
as mobile devices with Internet access (such as cell
phones) become more prevalent. Herzberg [7] pro-
vides an overview of challenges and opportunities
involved in using such devices for making secure
payments and other banking transactions. Two chal-
lenges and opportunities of a mobile payment system
are providing secure transactions and convenience. It
requires proper mechanisms for authentication and
authorization via modular design that can ensure a
minimal number of fraudulent transactions. In addi-
tion, the mobile payments provide the convenience
that stems from the use of mobile devices to initiate
transactions no matter where they are (i.e. ubiquity),
and from the uniform interface to accommodate mul-
tiple transactions.
Although m-payment systems have recently
gained a lot of attention [14], most payment protocols
proposed for wireless networks focus on the perfor-
mance and security issues [15-17, 21, 25]. Lee et al.
[17] enumerated the requirements and properties of
mobile commerce payment systems as confidential-
ity, authentication, integrity, authorization, avail-
ability, and non-repudiation. Kungpisdan et al. [16]
proposed a protocol for secure mobile payment which
used symmetric-key encryption. The symmetric tech-
nique was applied to reduce the computation of the
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
302 Journal of Information Technology and Architecture
parties involved and to satisfy the transaction secu-
rity properties. The proposed protocol offered the
ability to deal with failures and disputes among par-
ties. Kungpisdan et al. [15] also proposed a frame-
work for a practical mobile SET payment. Hwang et
al. [8] analyzed a previously-proposed m-payment
system and demonstrated its vulnerability. Cimato
[5] proposed an authentication protocol for GSM Jav-
acards. Popescu and Kanpskog [21] proposed an
anonymous mobile payment system that prevents
blackmailing based on the secure group blind sig-
nature scheme. Their biometrics-based system en-
abled the user of mobile devices to use fingerprint,
voice, or iris authentication, and other biometrics
technologies for secure online payments.
The existing mobile payment systems support
either prepaid, debit, or credit cards. As the name
implies, the prepaid method uses a SIM card that
stores electronic money or a virtual bank account
that can be accessed through message exchange sys-
tems such as SMS. The debit payment option is
made through ATM cards issued by the customer’s
own bank. It securely moves checking account funds
without the hassles or risks involved with paper
checks and without finance charges associated with
credit cards. Security is provided mostly with per-
sonal identification numbers (PIN). It is widely used
in Europe in retail businesses and it is predicted to
be even more widely used in the future.
Credit card payments are the most prevalent
method in e-commerce at present. The credit card
payment process is known to work almost anywhere
and people trust their trade labels like Visa or Mas-
terCard. Credit accounts are easy to implement
because they are less time critical: the details of pay-
ments made during daytime can be reported to the
credit card company as a batch job during the night
or according to the agreement between the merchant
and the credit card company. It is plausible to
assume that the people who use credit cards today
will use them as a payment method in m-commerce.
It is widely assumed that mobile devices will be inte-
grated with credit card payment systems [27], and in
this paper, we focus on extending the functionality of
the existing credit-card payment mechanism.
Much research on credit card payments for e-com-
merce focuses on security issues, such as fraud pre-
vention and reduction [24]. In order to reduce fraud,
the card issuing banks, such as American Express,
Discover, and MBNA, may issue a one-time tempo-
rary credit-card number (instead of a permanent
card number). A single-use credit card system was
studied by [22].
There have been studies of divisible e-cash pay-
ment protocols [3, 19]. These studies focus on pay-
ment solutions that ensure anonymity while allow-
ing electronic coins to be divisible. That is, each coin
can be spent incrementally as long as the total pur-
chases are limited to the monetary value of the coin.
These approaches deal with multiple purchases and
multiple merchants, while this paper is about one
purchase with one merchant but with multiple credit
card issuers.
Few studies on credit card payments focus on the
user-side issues. The customer may face a complex
utility optimization problem during each purchase.
The customer may need to determine which card or
which set of cards would be the best to use among
multiple cards for this particular purchase. The
user’s decision problem is to select a credit card or
a combination of cards to maximize her benefits
based on card features, such as incentives, interest
rates, etc. The solution of this decision problem will
be a recommendation of the best combination of
credit cards, so it can be viewed as a special kind of
a recommender system. While most of the early rec-
ommender systems are based on simple database
queries, our system is based on fuzzy logic [23].
Fuzzy logic has been used in many e-commerce
applications. It was used to handle the uncertainty
associated with customer choice when modeling
market structure [9]. The potential of a visitor as a
purchaser of services was classified using fuzzy
methods based on her demographic information [28].
Fuzzy logic was used to cluster customers into
several groups, based on their data [29]. Miao et al.
[18] conducted a case study on a recommender system
supporting automobile purchases. While these studies
discussed general and widely accepted practices of
fuzzy technology for e-commerce, we have used fuzzy
Flexible Payment Recommender System
Journal of Information Technology and Architecture 303
logic to model the customer’s perspective of credit-
card usage and solve the payment decision problem
based on the customer’s preferences.
Fuzzy technologies have also been applied to solve
optimization problems such as electronic production
assembly [11] and job grouping [9]. Similar to these
studies, we have applied the fuzzy technique to our
optimization problem. We believe that the card-
selection criteria, modeled by fuzzy sets, follow the
user’s intuitive requirements well. To model multiple
criteria, we have adapted the aspects of multi-objec-
tive optimization [6, 13]. Each criterion is represent-
ed by a fuzzy set and aggregated into an objective
function [11]. The user preferences for criteria are
flexibly weighted, such that the preferred criteria
will more significantly affect the output of an objec-
tive function [9]. Then, all the weighted criteria are
summed up to compute a single value and the best
solution is recommended based on that value [6, 13].
We have adapted this traditional approach to deal
with a variety of card features.
A preliminary survey of the user acceptance of
mobile payment decision support systems indicated
that card management is strongly favored, in addi-
tion to cost, security and convenience factors [25].
This paper proposes a “flexible payment recom-
mender system.” The proposed system recommends
the best combination of credit cards to use that is
consistent with the customer’s preferences, while
simplifying the potentially complex process of choos-
ing the card combination when many card features
and policies are to be considered.
A customer’s preferences and spending habits will
affect the use of a particular card. This has led us to
consider the personalization and customization
issues. Acquiring user preferences by asking them,
however, should be done carefully because of the phe-
nomenon, called “paradox of the active user” [2]. It
implies that a user is irrational in that she does not
want to spend initial time to set up the system envi-
ronment although it will bring some benefits in the
long term. Therefore, a system should be equipped
with reasonable default values for its parameters
and do its computations with a minimum of user
inputs, if possible. In our system, the initial default
values are obtained through statistical analysis of an
off-line user survey described in Section 6.
Apart from a user’s self-profiling, the prevailing
methods used for personalization are based on time-
series historical records (e.g., the items purchased
previously by the customer are used in Amazon.
com’s personalization) and/or horizontal survey
records (i.e., purchasing patterns of millions of buy-
ers adapted in a marketing strategy). Both methods
refrain from requiring the user to do extra work for
setting up her preferences, but implement a system
to learn users’ preferences and/or users’ behavioral
patterns. In our system, the survey method has been
used for estimating a typical user’s preferences, and
fuzzy sets have been formulated for user personal-
ization.
3. The Divisible Card Payment Decision
Model
The credit card(s) selection involves decision mak-
ing with multi-criteria, and user’s evaluation or pref-
erence on one criterion over another criterion may
not be always precise. In addition, each person may
exhibit different decision behaviors in different con-
texts. The same person may evaluate the same set
of criteria in evaluating the credit cards based on the
situation or context one is in. For example, one will
make a different decision on the priority of credit
card criteria when he/she is under tight financial
constraint and when one is under normal situation
with no additional stress, etc. For instance, one
would rate the interest rate more important than the
promotional incentives in the first case, while in the
second situation, one may be more interested in the
promotional incentives. This situational variation
needs to be captured in the decision model. Thus, the
traditional decision models based on precise values
or weights in one normal circumstance may not cap-
ture the real complex decision making situations. In
other words, one needs to decide the situation assess-
ment of oneself, and then prioritize the evaluative
criteria accordingly. In order to capture this decision
making behavior in different situations, we use a
pair-wise comparison of the situations, and find dif-
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
304 Journal of Information Technology and Architecture
ferent weights for each situation. This is similar to
a group decision making even though only one per-
son is involved in evaluating the criteria in judging
the cards. It is as if there is different persona of one
person and each persona influences differently in
decision making. The priority vector (weights) of dif-
ferent persona will be computed in our model.
Often users use evaluative or subjective linguistic
expressions to express the card preferences or the
criteria preferences. For instance, Amex card A is
rated “very good” while Master card M is “good”, or
the interest rates can be rated “very important” in a
situation A, while the promotional incentive factor is
“low in importance” in the same situation, and so on.
To reflect this imprecise nature of user’s preference
criteria in credit card rankings and payment divi-
sions, we model the credit card payment decision
problem with a fuzzy set theory combined with the
Analytical Hierarchy Process (AHP) method.
To solve the divisible payment decision problem,
we adopt fuzzy set theory and user-chosen weights
and thresholds, as described below.
Given a set of credit cards C={c1, c2, cn}, and a
set of policies or criteria applicable to each card,
P={p1, p2, pm}, we define the customer’s preferences
for different policies as weights dependent on the pol-
icies. More specifically, we define the customer’s
preference for policy j on card i as wi(pj), where
, , and .
At present, we assume the customer does not
distinguish between the same policies on different
cards. That is, the same weights are used for the
same policies on all the n credit cards (i.e.,
=…= ). We use w(pj) instead of
wi(pj). For instance, if a customer assigns a higher
weight to the interest rate and a lower weight to the
frequent flyer miles, she is assumed to have the same
weights for the interest rate and frequent flyer miles
policies on all of her cards. Of course, the preferences
of different policies and criteria vary among custo-
mers.
The customer may have a preferred value for each
policy. For example, an interest rate below the prime
rate or 1,000 frequent flyer miles might be viewed as
favorable. This is modeled by a threshold value of
policy j, t(pj). Similar to the case with weights, we use
the same threshold value for the same policy on
different cards. Each policy j on each credit card i is
viewed to have a value, vi(pj). When this value is
below the threshold value, the degrees of goodness of
the policy can be measured. The degree of goodness
of a policy is computed through fuzzy membership
values. Equation (1) represents the fuzzy value for
policy j on card i, fi(pj). The fuzzy membership value
is “1” if the corresponding value of the policy/criterion
is equal to or greater than the threshold value.
Otherwise, the membership value will be calculated
by the first part of equation (1).
fj(pj)= (1)
The fuzzy membership values are determined for
0 wi pj( ) 1≤ ≤ 1 i n≤ ≤ 1 j m≤ ≤
w1 pj( )=w2 pj( ) wn pj( )
vi pj( )t pj( )------------- , if vi pj( )<t pj( )
1 , otherwise⎩⎪⎨⎪⎧
Figure 1. Examples of fuzzy membership functions
Flexible Payment Recommender System
Journal of Information Technology and Architecture 305
each criterion or policy of each credit card separately.
There are two types of criteria. First, the larger the
value of the criterion is, the more desired it is by the
user, as shown in Figure 1(a). Second, the smaller
the value of the criterion is, the more desired it is,
as shown in Figure 2(b). To reduce the complexity of
the user interface and to enhance user friendliness
for the population that is not familiar with fuzzy sets,
we use the membership functions1 as shown in
Figure 3. Figure 1(a) and Figure 1(b) show the
shapes of a right and a left shoulder, respectively.
More complex user-defined fuzzy functions can be
used at the expense of additional computing over-
head. Note that the fuzzy membership function is not
generated until the user determines the threshold
value.
Figure 2 shows a fuzzy membership function for a
policy/criterion of the closing date. Note that the
criterion used is the number of days remaining until
the closing date (i.e., the difference between the
closing date and the current date). Most people
would prefer to use a credit card with the closing
date further away. In Figure 2, “10” is used as the
threshold value, and the fuzzy membership values
for the remaining days of 4, 6, 8, 13, and 18 are 0.4,
0.6, 0.8, 1, and 1, respectively. For some credit card
usage policies, such as “the credit card can only be
used to purchase office supplies, the criterion is
binary, and the fuzzy membership function gener-
ates either 1 or 0 as shown in Figure 3.
Using the membership function and the preference
weights for each policy, the PR agent computes the
optimal card combination through the following
steps. First, the weights and fuzzy sets for credit card
policies and criteria are initialized based on the
customer’s preferences. The criteria modeled in this
paper include the amount of available credit before
reaching the limit, the interest rate, the closing date,
and the bonus rate. Second, the fuzzy membership
value of each criterion of each card is determined by
the corresponding fuzzy set. Third, the PR adds up
the weighted fuzzy membership values of the criteria
of each card and generates a payment suggestion
based on the card with the maximum value. Finally,
the PR proposes to the user to pay the purchase
amount with the selected credit card. If the selected
credit card does not have enough available credit for
the whole payment, the PR repeatedly performs thehttp:// www.research.ibm.com/able/doc/reference/com/ibm/
able/rules/doc-files/arlDefSets.html
Figure 2. A fuzzy membership function for the closing date
Figure 3. A fuzzy membership function for a binary
criterion
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
306 Journal of Information Technology and Architecture
above procedure until a combination of credit cards
covers the total purchase price.
The payment recommender algorithm is shown in
Algorithm 1. Let n and m be the number of credit
cards and the number of policies, respectively. Let ci
be credit card i. Let fi(pj) and w(pj) be the fuzzy set
membership value and the weight of policy j on card
i, respectively. Let Evali be the evaluation value
summing up all the weighted fuzzy membership
values of card i. Let CT be the card selected for
payment at the T-th iteration and Totalamount be
the total purchase amount. Let Availi and Payi be the
available credit and the payment of card i, respec-
tively.
Algorithm 1: Fuzzy Payment Recommender Algo-
rithm
1. Define the customer’s preferences by initializing
the weight, w(pj), and the threshold value, t(pj),
for each criterion/policy.
2. Calculate the fuzzy membership value, fi(pj), for
each criterion/policy on each credit card, where
, and .
3. Obtain the evaluation value, Evali for each credit
card ci by adding up the fuzzy membership values.
(2)
4. Determine the best card among those with
remaining credit limit by finding the one with the
highest evaluation value.
CT = (Evali), where AvailT > 0 (3)
5. If , pay the whole amount
with the best card T and adjust the following:
AvailT= AvailT − Totalamount, PayT= Totalamount,
and Totalamount=0.
Otherwise, make a partial payment with the best
card selected and adjust the following:
PayT= AvailT, AvailT= 0, and Totalamount =
Totalamount −AvailT
6. If Totalamount = 0, recommend the list of cards,
{Ci}, with the suggested payment amount for each
card. Stop.
Otherwise, goto step 3.
4. The Flexible Card Payment Processing
Infrastructure
In order to handle the proposed payment with
multiple cards the existing payment processing
infrastructure needs to be extended [20]. The current
infrastructure for the credit card payment processing
is depicted in Figure 4 and described as follows.2 The
customer finds the desired products from the mer-
chant’s Web site (Step 1); When she is ready to make
a payment, the customer enters the billing informa-
tion on the merchant’s secure payment gateway
(Step 2); The payment gateway provides the billing
information (such as credit card number, card-
holder’s name, billing address, expiration date, etc.)
to the merchant account that has been set up by the
online merchant account provider (Step 3). The mer-
chant account provider offers the credit card pro-
cessing service for each merchant account. It trans-
fers the billing information to the issuing bank of the
customer’s credit card for authorization (Step 4). The
issuing bank checks if the credit card information is
valid and if the credit card has a sufficient balance
to cover the purchase. If so, it sets aside the amount
of purchase from the customer’s account for the mer-
chant. If the credit card is invalid or the credit limit
has been reached, the issuing bank sends back a
denial code to the merchant account (Step 5). The
approval (or denial) code is passed to the payment
gateway back from the secure merchant site (Step 6).
The approval code is passed to the customer. Usu-
ally, the payment gateway emails the customer the
payment receipt (Step 7). At the end of the day, the
merchant requests to settle all the transactions of
the day. The merchant account provider offers the
merchant account settling service. It sends the
request to capture funds to the acquiring bank (Step
8). The acquiring bank forwards the request to the
issuing bank (Step 9). The card issuing bank pays
1 i n≤ ≤ 1 j m≤ ≤
Evali=
j 1=
m
∑ fi pj( ) w pj( )×
Maxi 1=
n
AvailT Totalamount>
See http://www.electronictransfer.com/ for more details.
Flexible Payment Recommender System
Journal of Information Technology and Architecture 307
funds to the acquiring bank and the funds are depos-
ited to the merchant’s bank account. The actual
funds reach the merchant’s checking account in
approximately two business days (Step 10).
To support flexible card payments, the existing
infrastructure needs two modifications. First, a soft-
ware agent called Payment Recommender (PR) is
added to the client side. For each purchase, the PR
recommends to the customer the optimal combina-
tion of credit cards to use. If the customer accepts its
suggestion, the PR generates the Virtual card (V-
card in short) which is a set of recommended cards
and their respective payment amounts. The PR gen-
erates a unique V-card number, the total amount in
the card, the timestamp, and the billing information
of all the recommended cards. As the effectiveness of
the PR affects the performance of the whole payment
system, developing an effective and efficient PR that
can accommodate numerous card features and pol-
icies is an important research issue. We discuss a
prototype payment recommender, called fuzzy pay-
ment recommender (fuzzy PR), in Section 5.
Second, special software called a Virtual Card
Manager (VCM) is added at the merchant side to
handle the V-card payment request. When the V-
card is up for approval, the VCM decrypts the divis-
ible payment information (i.e., which cards and what
amounts should be charged for the V-card purchase),
and forwards the billing information to each card
issuer involved in the V-card. Unlike the current pro-
tocol that contacts one credit card issuer for ap-
proval, the VCM needs to communicate with all the
issuing banks mentioned in the V-card. The VCM
could process each card approval sequentially, one at
a time, or request the approval of multiple cards con-
currently. Each card-issuing bank sends back the
approval code to the VCM. When all the approval
codes have been received, VCM sends back the ap-
proval of the V-card to the payment gateway.
Figure 5 depicts the proposed infrastructure for
the flexible card payment processing infrastructure.
First, the customer finds the desired products from
the merchant’s Web site. The Payment Recommender
agent computes the optimal card combination for the
payment and generates a V-card (Step 1). The V-card
information is passed to the payment gateway and
the V-card billing information is transferred to the
VCM of the merchant’s account. (Steps 2-3). The
VCM transfers the billing information for approval to
the issuing banks of each credit card used in the V-
card (Step 4). Each issuing bank checks whether the
credit card information is valid and whether the
credit card has sufficient funds. If so, it sets aside the
amount of purchase for the merchant. Each issuing
bank in the V-card sends back the approval (or
denial) code to the merchant’s VCM (Step 5). The
VCM waits until all pertinent card issuers send back
their approval (or denial) codes. When the approval
codes from all card issuers in the V-card are avail-
able, the VCM generates an approval code for the V-
card, and forwards the code to the payment gateway
(Step 6). The approval code is passed to the customer.
Figure 4. The Existing Credit Card Payment Infrastructure
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
308 Journal of Information Technology and Architecture
The payment gateway emails the customer a pay-
ment receipt. The PR adjusts the credit card bal-
ances resulting from the current purchase with the
V-card (Step 7). At the end of the day, the merchant
requests to settle all the transactions of the day. Step
8 to 10 are the same as in the existing processing
infrastructure, namely, the merchant account sends
the request to capture funds to the acquiring bank.
The acquiring bank forwards the request to the
issuing banks. The card issuing banks pay funds to
the acquiring bank and the funds are deposited to
the merchant’s bank account. The actual funds reach
the merchant’s checking account in approximately
two business days.
At Step 5, if any one of the issuing banks denies
the billing request, the V-card transaction is consid-
ered denied, and any approved requests should be
nullified. That is, the V-card approval request is han-
dled as an atomic transaction with multiple approval
requests. The V-card approval request, as an atomic
transaction, consists of distributed approval requests
(Tygar 1998) to each card issuer, similar to those in
a transaction processing system. For example, a V-
card transaction may consist of three approval re-
quests to the card issuers A, B and C, respectively.
As in a transactional system, either all of the re-
quests in the V-card need to be approved, or none of
them is considered approved. Suppose, A and B have
sent back approval codes, but C sent back a denial
code. Then the whole transaction (i.e. the approval
request to use the V-card) is considered denied, and
all other approvals are rolled back. This atomic prop-
erty of the V-card approval transaction ensures no
partial approvals or partial denials are possible. The
VCM takes care of this atomicity of the V-card’s
approval process.
As shown above, the flexible card payment archi-
tecture requires minimal modifications to the exist-
ing infrastructure, namely the PR (Step 1 above) and
VCM (Steps 4-6 above), while allowing flexible pay-
ment using multiple cards for a single purchase.
5. Payment Recommender Prototype System
We have developed the Web-based prototype of the
Fuzzy Payment Recommender system using JSP
and deployed it on an Apache Web server. The
prototype system provides two major functionalities
of PR: managing multiple cards and generating a
recommendation of cards for a purchase. The proto-
type also simulates the V-Card Manager that performs
the multiple credit card approval process.
5.1 Card management Component
This component allows the customer to register
and log in. When the customer logs into the PR
system, the main menu comes up with three sub-
menus: ‘My Card List’, ‘My Preference’, and ‘Create
A V-Card’. The ‘My Card List’ allows the customer
to view the list of all her cards, add an additional
card, to specify card features, such as interest rate,
Figure 5. The Flexible Card Payment Infrastructure
Flexible Payment Recommender System
Journal of Information Technology and Architecture 309
balance, etc. It is intended to connect to the card
provider’s server to automatically receive these
properties, such as the existing balance information
(and therefore the currently available credit). This is
similar to the feature of automatically receiving the
existing transactions as used in commercial personal
finance software, such as Quicken and Money. In the
prototype, preferences and features of cards are
captured through a series of simple questions.
5.2 Payment Recommendation Component
This component generates a virtual card, i.e. one
or more cards to be used for the payment. When the
customer wants to purchase a product, she enters the
purchase amount and pushes the “Recommend cards”
button. The PR performs the fuzzy membership-
based optimization described in Section 3. For ex-
ample, the customer enters the purchase amount of
$500. This information, in conjunction with the pre-
viously entered information about credit card fea-
tures and usage preferences, is used to perform the
optimization. Figure 6 shows that the PR considers
the card features such as closing date, interest rate
and bonus rate, and shows the fuzzy membership
value of each feature for each card. The preference
weights based on the user’s card usage preferences
are also used.
The PR system generates a list of cards to be used
and the amount of charge against each card, creates
a virtual card (V-Card), and logs its information,
such as transaction number and amount, timestamp,
expiration date and current status.
5.3 Virtual Card Processing Component
With the user’s confirmation, the V-Card infor-
mation is sent to the merchant bank where the
Virtual Card Manager (VCM) handles the V-card
approval process with each card’s issuing bank. We
have implemented a VCM component with simula-
tions of three credit card issuing banks. For simpli-
city, each credit card bank has a back-end database
with balance information. When the simulated card
issuing bank receives the approval request from the
VCM, its server authenticates the account infor-
mation and checks the available credit line in the
database to see whether this transaction amount is
valid. After validation, each issuing bank returns
either a denial message or an approval code back to
Figure 6. The Fuzzy Payment Optimization
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
310 Journal of Information Technology and Architecture
the VCM. The approval/denial code is then sent back
to the PR system, which adjusts the balance amount
of each card involved in the particular transaction.
After a V-card purchase has been approved, the
transaction information is recorded into the database
of the merchant site. At the end of the day, the
merchant can log into the acquiring bank and re-
quests to settle all the transactions of the day. The
merchant account sends the requests to each involv-
ed card issuing bank to capture funds, and the funds
are deposited into each acquiring bank. The issuing
banks update their own databases by reducing the
available credit. For the acquiring banks, the update
increases the current balance.
5.4 Implementation Options
The Payment Recommender system can be imple-
mented on Electronic commerce sites or on a smart
phone. Recent mobile payment infrastructure by
Google, called “Google Wallet”, allows credit card
payments using smart phones and Near Field Com-
munication wireless network [30]. However, Google
Wallet is not allowing multiple card payments for a
single purchase. Our PR system can be implemented
as a mobile app for smart phones, and use the exist-
ing payment infrastructure, like Google Wallet,
using one credit card as the only option. Or it can
extend the mobile payment infrastructure to connect
to the several credit card companies for a payment,
when multiple card payment option is turned on. The
option of having the end users to configure the loy-
alty and discount benefits, not just from Google pro-
viders, and automatically recommending credit cards
(one or more) can add values for the customers’ pur-
chase experience.
6. Survey Studies on Payment Recommen-
der System
To find out the general public’s perception and
acceptance of the PR system and to find out the
major factors in card payment decisions, we con-
ducted a preliminary survey on 68 subjects with 24
females, 43 males and 1 unspecified. The survey
questionnaire was posted on the Web and the sta-
tistical data analysis was performed using the SAS
package.
6.1 General Acceptability Analysis
First, we investigated the correlations among var-
ious variables to see whether there exist linear rela-
tionships between them. In the correlation matrix
(see Appendix A), the Pearson correlation coefficients
indicate the strength of the linear relationship bet-
ween two variables. The matrix shows that the cor-
relation between “age (AE)” and “number of cards a
user has (NC)” is 0.58 at the significance level 0.01.
There is a positive linear relationship between “age
(AE)” and “income (IE)” at the significance level of
0.01. The variable, “years a user has used a computer
(YC)” has a positive relationship with “years a user
has used the Web (YW)” and “the number of cards
a user owns (NC)” at the significance level of 0.01.
We conducted the correlation analysis on three
variables: “willingness to use a mechanism which
allows several cards for single online purchase (WM)”,
“willingness to use computer software to manage
credit cards (WP)” and “willingness to use a personal
card payment recommender software with automatic
filling of payment information (WA)”. The three vari-
ables, WM, WP and WA, are considered to be the
indicators of the usability and the user acceptance of
the flexible payment system.
The value of the correlation coefficient between
WM and WP is 0.49 and between WP and WA is
0.58, both of which indicate positive relationships at
the significance level of 0.01. The correlation coeffi-
cient between WM and WA is 0.30, which indicates
a weak relationship between the two at the signif-
icance level of 0.05.
Only the “feeling of comfort with credit card(s)
(CC)” has minor relationships with WP and WA with
the coefficient values of 0.29 and 0.36, respectively,
at the significant level of 0.05. We could not find any
significant relationships between other variables
(e.g., age, income, etc.) and the three variables WM,
WP and WA. We concluded that a comfortable feel-
ing with a credit card(s) can potentially be a factor
Flexible Payment Recommender System
Journal of Information Technology and Architecture 311
that affects the usability and user acceptance of the
flexible payment system.
Secondly, we investigated the medians of WM, WP
and WA on a 5 Likert-scale as shown in Table 1. In
the 5 Likert-scale used in the survey, 4 meant “agree”
and 3 meant “neutral.” For example, for the question
“I would gladly use computer software to manage my
credit cards (i.e., Payment Recommender System
which recommends the best combination of available
cards for your online purchases),” subjects answered
“agree” on average. We note that each of the three
usability indicator variables forms a normal distri-
bution according to Kolmogorov-Smirnov analysis.
Subjects showed positive answers with mean 3.44 in
using a divisible card payment system (i.e.,WM) and
in using computer software to mange credit cards
(i.e., WP) with mean 3.41. Compared with WM and
WP, on the other hand, subjects showed less interest
in auto filling of payment information (WA) with
mean 3.05.
Finally, we investigated potential concerns about
the flexible payment system. Two open-ended ques-
tions were given: (a) What feature is most important
for you in a personal card management system? and
(b) What would stop you from using a personal card
management system? Figure 7(a) summarizes the
answers to the question (a), and Figure 7(b) sum-
marizes the answers to the question (b). As shown in
Figure 7, the public is most concerned with security.
Other features such as convenience, alerts, consoli-
dation, privacy, incentives like discounts, speed, and
user friendliness are also found to be desirable for a
good payment recommender system.
6.2 Factor analysis to identify the underlying
major factors
We applied factor analysis to identify major factors
and to evaluate the relative importance among the
identified factors. First, various criteria were re-
duced to several main factors which explain the vari-
ances in the data points of the user answers. Second,
the final commonality estimates derived from factor
analysis were applied to evaluate the relative impor-
tance among the factors. Finally, the importance of
each criterion for its corresponding factor was ana-
lyzed. This information is used as the weight in the
weighted-sum approach of multi-criteria optimiza-
tion in an example.
(1) Factor analysis for classifying card selec-
tion criteria: Given the question in our survey,
“The following are important when choosing one
credit card over another. What do you think?” with
the specification of 12 criteria used in choosing one
card over another (e.g., frequent flyer miles, etc.),
subjects marked their answers on the 5 Likert scale.
To figure out the underlying important factors (prin-
ciples) over multiple criteria, we used factor analysis.
Our assumption was that a variety of criteria could
be reduced to a few important factors which effec-
tively explain the variances of user answers on cri-
teria and rank each criterion within its correspond-
Figure 7. Users’ concerns about a payment recommender system
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
312 Journal of Information Technology and Architecture
ing factor.
Table 1 shows rotated factor loadings. We obtained
3 factors from 12 criteria.
The numbers in Table 1 represent the correlations
between three factors and the criteria after rotation.
Shaded boxes show strong correlations of each factor
corresponding criterion. The correlation between
each criterion and its corresponding factor contrib-
utes to computing the weight of a criterion. “Cash
Back” is most correlated with factor 1 and “Discount
Rate” is least correlated with factor 1. “Interest
Rates” is most correlated with factor 2 followed by
“Annual Fee” and “Available Balance”. “Closing
Date” is most correlated with factor 3 followed by
“Card Brand.”
(2) The importance of each factor in the total
commonality estimates: By observing the strength
of correlations between the factors and the corre-
sponding features, we realized that the first factor
deals with policies/criteria that provide potential
incentives to users. We named this factor Incentives.
The second factor deals with numeric features of a
card, so we named it Costs. The third factor is named
Perception, which deals with nonnumeric features
such as a card brand or reputation. The variances
explained by factor 1, 2 and 3 were 3.692, 1.529 and
0.93, respectively. The total commonality estimate
(i.e., the total estimate of the proportion of common
variance in a variable) was 6.153. Figure 8 shows the
break down of the variance. The first factor, Incen-
tives, accounts for most of the variance, followed by
Costs and Perception.
6.3 An example of evaluating a credit card using
the identified factors
Based on the survey analysis, credit card users
seem to prefer a card that provides a feature(s) called
Incentives. Among the policies/criteria belonging to
this group, “Cash Back” is the most important to
credit card users. Therefore, “Cash Back” which
belongs to factor 1 has a higher weight than any
other feature. On the other hand, “Interest Rate”
which belongs to Costs is another considerably
important feature.
The formula shown in the step 3 of the Fuzzy Pay-
ment Recommender Algorithm 1 is illustrated as fol-
lows:
Evali=
, where l is the number of factors.
The range of i is , and the range of j is
, where n is the number of credit cards a
user has and m is the number of credit card features/
criteria classified with each factor. In the survey
above, m = where m1=7, m2=3, and m3=2, as
shown in Table 1; ck is the proportion of the total
commonality estimates explained by the k-th factor,
which was computed in Figure 8; is the fuzzy
membership value of the j-th criterion which belongs
to the k-th factor; is the importance (weight)
of the j-th criterion which belongs to the k-th factor.
Note that the weight is the correlation between the
k 1=
l
∑ ck
j 1=
mk
∑ fi pjk( )*w pj
k( )⎝ ⎠⎜ ⎟⎛ ⎞
×⎝ ⎠⎜ ⎟⎛ ⎞
=
k 1=
l
∑
j 1=
mk
∑ ck* pj
k( )
*w pjk( )
1 i n≤ ≤
1 j m≤ ≤
k 1=
l
∑
fi pjk( )
w pjk( )
Table 1. Rotated Factor Loadings
Factor: 1
(Incentives)
Factor: 2
(Costs)
Factor: 3
(Perception)
Cash Back 0.78628 0.08356 -0.04894
Shopping Benefits 0.72902 -0.04621 0.25708
Fraud Insurance 0.71901 0.04097 0.10903
Award Package 0.70659 0.20399 0.08128
Purchase Protection 0.68083 0.04002 0.06552
Frequent Flyer Miles 0.66427 0.04594 0.23345
Discount Rate 0.63552 0.28317 -0.00643
Interest Rates -0.07788 0.89948 0.19549
Annual Fee 0.03169 0.63134 -0.09360
Available Balance 0.12489 0.32993 0.25644
Closing Date 0.08016 0.10106 0.69207
Card Brand (e.g. VISA) 0.11255 0.00253 0.66072
Figure 8. Variances explained by three factors
Flexible Payment Recommender System
Journal of Information Technology and Architecture 313
j-th criterion and the k-th factor, shown in shaded
areas in Table 1.
An example of evaluating credit card i with three
factors (i.e., l = 3) is as follows.
Evali =
Let’s focus on the third term, .
where = Card Brand (e.g. VISA), =Closing
Date, =0.66072, =0.69207 from Table 1
and c3=0.15, the perception factor, from Figure 8.
Assuming the fuzzy membership values of
=0.5 and =0.7, we can compute the third term
as follows.
.
That is, the value of the third term of the evaluation
function, Evali, is 0.12. Once the first and second
terms are computed, a value for the evaluation of the
i-th card is obtained. After that, computation with
the formula in step 4 of Algorithm 1 is performed to
eventually find the best combination of credit cards.
7. Conclusions and Future Work
This paper presents the flexible card payment
infrastructure for mobile commerce that recom-
mends one or more credit cards for acustomer. The
recommender system can suggest multiple cards to
divide the payment of a single purchase. The pro-
posed payment system uses the client Payment Rec-
ommender (PR) agent and the merchant’s card
approval management software called Virtual Card
Manager (VCM). The major benefits of our proposed
system include the following. First, the proposed
solution requires minimal modifications to the exist-
ing online and mobile payment infrastructure. Sec-
ond, the fuzzy set based payment recommender (PR)
agent adds value to the credit-card users by custom-
izing each payment according to the user’s prefer-
ences. The use of weighted criteria in selecting a card
over another allows users to express relative pref-
erences between criteria. The use of fuzzy sets allows
time-dependent and parameter-based computations
which determine the evaluation of a card on a daily
basis.
We provided a detailed Payment Recommender
algorithm and developed a prototype system to dem-
onstrate the functionalities of the proposed PR agent.
We also conducted a user-acceptability survey of the
proposed flexible payment approach and of the user
preference criteria in card selection. The factor anal-
ysis of different criteria was performed to compute
the relative importance of the each card selection cri-
terion and an example was provided to illustrate how
the differences in a user’s preferences are reflected
in the algorithm in terms of different weights.
Future research issues include extending the rec-
ommender system to utilize the user’s transaction
and payment histories to derive the credit card
related preferences, identifying security challenges
in mobile payment with the new infrastructure, and
finding a feasible business model to minimize the
costs involved in the merchants server side trans-
actions generated by using multiple cards.
References
[1] Barnes, S. “The Mobile Commerce Value Chain: Anal-
ysis and Future Developments”, International Journal
of Information Managemet, Vol. 22, pp. 91-108, 2002.
[2] Carroll, J.M. and M.B. Rosson, “The Paradox of the
Active User”. In J.M. Carroll (Ed.), Interfacing Thought:
Cognitive Aspects of Human-Computer Interaction.
Cambridge, MA: MIT Press, 1987.
[3] Chan, A.,Y. Frankel and Y. Tsiounis, “Easy Come
Easy Go Divisible Cash”, Proceedings of Eurocrypt '98
(Lecture Notes in Computer Science). Springer-Verlag,
1998. Available at http://www.ccs.neu.edu/home/yian-
nis/pubs.html.
k 1=
l
∑
j 1=
mk
∑ ck*fi pj
k( )*w pjk( )=
j 1=
m1
∑ c1*fi pj
1( )*w pj1( )
+
j 1=
m2
∑ c2*fi pj
2( )*w pj2( )=
j 1=
m3
∑ c3*fi pj
3( )*w pj3( )
j 1=
m3
∑ c3*fi pj
3( )*w pj3( )
p1
3p2
3
w p1
3( ) w p2
3( )
f1 p1
3( )
f2 p2
3( )
j 1=
m3
∑ c3*fi pj
3( )*w pj3( )=0.15×
j 1=
m3
∑ fi pj3( )*w pj
3( )⎝ ⎠⎜ ⎟⎛ ⎞
=0.15×
j 1=
2
∑ fi pj3( )*w pj
3( )⎝ ⎠⎜ ⎟⎛ ⎞
=0.15× f1 p1
3( )*w p1
3( )+f1 p2
3( )*w p2
3( )( )
=0.15*f1 p1
3( )*w p1
3( )+0.15*f1 p2
3( )*w p2
3( )
=0.15*0.66072*0.5+0.15*0.69207*0.7=0.12
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
314 Journal of Information Technology and Architecture
[4] Chun, S., Y. An, J. Geller and S. Park, Fuzzy Virtual
Card Agent for Customizing Divisible Card Payments,
K. Bauknecht, B. Proll and H. Werthner (eds.) Lecture
Notes in Computer Science 3590, E-Commerce and
Web Technologies, the 6th International Conference
(EC-Web 05), Springer, 2005, pp 287-296.
[5] Cimato, S. “ Design of Authentication Protocol for
GSM Javacards”, LNCS Vol. 2288, 2002.
[6] Fuller, R and C. Carlsson, “Fuzzy Multiple Criteria
Decision Making: Recent Developments”, Fuzzy Sets
and Systems, Vol. 78, No. 2, pp. 139–153, 1996.
[7] Herzberg, A. “Payments and banking with mobile per-
sonal devices”. Communications of the ACM, Vol 46,
No 5, May 2003.
[8] Hwang, S. D. Lee, D. Han and J. Ryou. “Vulnerability
of a Mobile Payment System.” International Workshop
on Information Security Application, 2004.
[9] Jain, V. and R. Krishnapuram, “Applications of Fuzzy
Sets in Personalization for e-Commerce”, Proceedings
of IFSA-NAFIPS 2001 Conference, July 2001.
[10] Johtela, T, J. Smed, M. Johnsson, and O. Nevalainen,
“Fuzzy Approach for Modeling Multiple Criteria in
The Job Grouping Problem”, In M. I. Dessoyky, S. M.
Waly, and M. S. Eid, editors, Proceedings of the 25th
International Conference on Computers & Industrial
Engineering, pp. 447-450, New Orleans, LA, Mar,
1999.
[11] Johtela, T, J. Smed, M. Johnsson, and O. Nevalainen,
“Single Machine Scheduling with Fuzzy Multiple Cri-
teria Optimization in Electronic Assembly”, in Ylini-
emi and Juuso (eds.), Proceedings of TOOLMET'98
Symposium, pp. 121-126, Oulu, Finland, 1998.
[12] Jupiter Research Corporation, Mobile Commerce in
Europe, Marketresearch.com, August, 2004.
[13] Kim, I. Y. and de Weck, O, “Adaptive Weighted Sum
Method for Bi-objective Optimization”, Structural and
Multidisciplinary Optimization, Vol. 29, pp. 149-158,
2005.
[14] Krueger, M. “The future of M-payments-Business
Option and Policy Issues” Background Paper No.2.,
Electronic payment Systems Observatory (ePSO) August
2001. Available at http://epso.intrasoft.lu/papers/Back-
grnd-2.pdf
[15] Kungpisdan, S. B. Srinivasa, P. D. Le, “A Practical
Framework for Mobile SET Payment”, International e-
Society Conference, 2003.
[16] Kungpisdan, S. B. Srinivasa, P. D. Le. “A Secure
Account-based Mobile Payment Protocol”, IEEE Inter-
national Conference on Information Technology: Coding
and Computing. 2004.
[17] Lee, C. W. Kou, and W. Hu. “Mobile Commerce Secu-
rity and Payment Methods”, Advances in Security and
Payment Methods for Mobile Commerce. Idea Group,
Inc. 2004.
[18] Miao, C., Q. Yang, H. Fang, and A. Goh,. “Fuzzy Cog-
nitive Agents for Personal Recommendation”, Proceed-
ings of the 3rd International Conference on Web Infor-
mation Systems Engineering, 2002.
[19] Nakanishi, T. and Y. Sugiyama. “Unlinkable Divisible
Electronic Cash,” Proceedings of Third International
Information Security Workshop, ISW 2000, Australia,
2000, Lecture Notes in Computer Science 1975, pp.
121-134, Springer, 2000.
[20] Park, S., Chun, S. and Cho, J., “An Infrastructure for
Customizable and Divisible Card Payments for Online
Purchases,” Proceedings of the 35th Annual Meeting of
the Decision Sciences Institute (DSI 2004), Boston, MA,
2004, pp 5091-5096.
[21] Popescu, C. S. Knapskog. “An Anonymous Mobile
Payment System based on a Group Blind Signature
Scheme”. Proceedings of the IST Mobile & Wireless
Communications Summit 2004, June 2004.
[22] Rubin, A. D. and R. N. Wright, “Off-Line Generation
of Limited-Use Credit Card Numbers.Financial Cryp-
tography, pp. 196-209, 2001.
[23] Schafer, J. B., J. A. Konstan and J. Riedl, “E-commerce
Recommendation Applications”, Data Mining and
Knowledge Discovery. Vol.5, pp115-152, 2001.
[24] Shankar, U. and M. Walker, “A Survey of Security in
Online Credit Card Payments,” UC Berkeley Class
Notes, May, 2001.
[25] Tech Republic, “Enabling Secure, Interoperable, and
User-Friendly Mobile Payments”, Mobile Payment
Forum White Paper 2002. Available at http://whitepa-
pers.techrepublic.com.com/whitepa-
per.aspx?docid=69050
[26] Tygar, J. D. “Atomicity versus Anonymity: Distributed
Transactions for Electronic Commerce,” Proceedings of
the 24th VLDB Conference, New York, USA 1998.
[27] WirelessDevNet.com, “Mobile Payments: Death of
Cash and the Credit Card?”, Available at http://
www.wirelessdevnet.com/news/2004/sep/24/
news3.html
[28] Yager, R., “Fuzzy Methods in E-Commerce”, Annual
Conference of the North American Fuzzy Information
Processing Society - NAFIPS 1999, p 5-11.
[29] Zhang, Y., M. Shteynberg, S.K. Parasad and R. Sun-
derraman, “Granular Fuzzy Web Intelligence Tech-
niques for Profitable Data Mining”, Proceedings of
2003 IEEE International Conference on Fuzzy Sys-
tems, pp. 1462-1464, May, 2003.
[30] Google, “Google Wallet” [computer software], 2011.
Flexible Payment Recommender System
Journal of Information Technology and Architecture 315
Available at http://www.google.com/wallet/
Appendix A– Correlation Matrix
Correlation analysis is used to find the linear rela-
tionship between two variables. We designed 14 vari-
ables for correlation analysis and abbreviation for
the variables is as follows:
AE : age, GR: Gender, IE: Income
YC : Years a user has used a computer (i.e., ques-
tion 4)
YW : Years a user has used Web (i.e., question 5)
FP : Frequency a user purchase product or services
on the Web (i.e., question 6)
NC : Number of cards a user owns (i.e., question 7)
AB : Average balance a user carries on each one of
his credit card (i.e., question 8)
FC : Frequency a user use a credit card for online
purchase (i.e., question 9)
SP : Average size of one of online purchases (i.e.,
question 10)
CC : Comfort to use credit card(s) (i.e., question 12)
WM: Willingness to use a mechanism which allow
several cards for single online purchase (i.e.,
question 14)
WP : Willingness to use computer software to man-
age credit cards (i.e., question 15)
WA : Willingness to use a Personal Card Manager
with auto filling payment information (i.e.,
question 16)
The table below is correlation matrix with the list of
variables. The numbers in the table are “Person Cor-
relation Coefficients” which indicate the strength of
the linear relationships between two variables. The
range of the numbers is from −1 to +1 where −1 indi-
cates a perfect negative correlation and +1 indicates
a perfect positive correlation. For example, the 0.58
on the position of NC (i.e., number of cards a user
owns) and AE (i.e., age) in the matrix says that the
tightness with which data points of NC and AE are
around the line is 58 percent. High correlation means
that the two variables are tend to form a linear rela-
tionship.
AE IE YC YW FP NC AB FC SP CC WM WP WA
AE 1 0.43 0.28 0.36 0.09 0.58* 0.23 -0.08 0.26 -0.05 -0.15 -0.06 -0.19
IE 0.43 1 -0.13 0.12 -0.26 0.15 0.11 -0.01 0.29 0.16 -0.00 -0.23 -0.05
YC 0.28 -0.13 1 0.42 0.06 0.40 -0.07 0.24 0.23 -0.09 -0.30 -0.28 -0.05
YW 0.36 0.12 0.42 1 -0.06 0.33 -0.21 0.00 0.12 -0.12 -0.15 0.03 -0.06
FP 0.09 -0.26 0.06 -0.06 1 0.07 0.00 0.33 0.03 -0.24 -0.18 -0.27 -0.09
NC 0.58* 0.15 0.40 0.33 0.07 1 0.01 0.06 0.04 0.04 -0.06 -0.09 0.00
AB 0.23 0.11 -0.07 -0.21 0.01 0.01 1 -0.19 0.04 -0.08 0.05 -0.12 -0.18
FC -0.08 -0.01 0.24 0.00 0.33 0.06 -0.19 1 -0.05 -0.07 0.08 0.21 0.10
SP 0.26 0.29 0.23 0.12 0.03 0.04 0.04 -0.05 1 0.09 -0.20 -0.00 -0.09
CC -0.15 0.16 -0.09 -0.12 -0.24 0.04 -0.08 -0.07 0.09 1 0.02 0.29 0.36
WM -0.15 -0.00 -0.30 -0.15 -0.18 -0.06 0.05 0.08 -0.20 0.02 1 0.49 0.30
WP -0.06 0.23 -0.23 0.03 -0.27 -0.09 -0.12 0.21 -0.00 0.29 0.49 1 0.58*
WA -0.19 -0.05 -0.05 -0.06 -0.01 0.00 -0.18 0.10 -0.09 0.36 0.30 0.58* 1
*indicates a relatively higher correlation.
Soon Ae Chun, Yoo Jung An, Sunju Park, June-Suh Cho and James Geller
316 Journal of Information Technology and Architecture
Soon Ae Chun is an Associate Professor in Business Department, and a member of the doc-
toral faculty of Computer Science Program, at the City University of New York, College of
Staten Island. Her research areas include database, information security and privacy, applied
informatics, such as Digital Government, e-business and e-Health; semantic approaches for
Workflow and Web services; Web 2.0 and mobile Web. She received her Ph.D.. in Information
Technology from Rutgers University, and M.S. in Computer Science from SUNY Buffalo. She
is a member of IEEE and the ACM. Contact her at [email protected]
Yoo Jung An Dr. Yoo Jung An received M.S. and Ph.D. degrees in Computer Science from
New Jersey Institute of Technology (NJIT) in January 2004 and 2008, respectively. Currently,
she is teaching IS and CS courses at Fairleigh Dickinson University (FDU) and NJIT as an
adjunct professor. Her research interests include Semantic Web, Health Informatics, Deep Web,
Data Warehouses and Artificial Intelligence.
Sunju Park is an Associate Professor in School of Business at Yonsei University in Korea.
Her research interests are the application of OR techniques to real-world problems, such as
online social networks, smart grid, healthcare, and telecommunications. She received her Ph.D.
in Computer Science and Engineering from University of Michigan, and M.S. and B.S. in Com-
puter Engineering from Seoul National University. Contact her at [email protected].
June-Suh Cho holds Ph.D. from Rutgers University, M.S. from New York University and B.A.
from KyungHee University. He is anassociate professor at Hankuk University's College of Busi-
ness Administration and Graduate School of Business. Dr. Cho had worked at several Invest-
ment Banks and Insurance Companies in New York City. He also had conducted research at
EC & Data Management group in IBM T.J. Watson Research Center. His research focuses on
e-Business, Multimedia Databases, Security, CRM, Mobile and Telecommunication technology.
He has published articles in IEEE, ACM, and other conferences and journals.
James Geller is a professor in the Department of Computer Science at the New Jersey Insti-
tute of Technology. His research areas include the structure and semantics of object-oriented
databases; Semantic Web search and mining, and the Unified Medical Language System, where
he and his colleagues have taken a leading role using regularities in the structure of medical
ontologies to assure their correctness. His distinguished publication record in major medical
informatics conferences and journals was crowned by co-editing the groundbreaking special
issue on Auditing of Terminologies of the Journal of Biomedical Informatics in 2009. He has
received numerous awards, including: the Excellence in Research Award for NJIT's College of
Computing Sciences (2010); NJIT Excellence in Teaching Awards (2003, 2011); and the College
of Computing Sciences Teaching Award (2003). He was named a Master Teacher in 2005. Geller
received his PhD and MS from the State University of New York at Buffalo in 1988.