You are on page 1of 7

International Journal of Advanced Computer Science, Vol. 3, No. 6, Pp. 267-273, Jun., 2013.

Manuscript
Received:
29,Mar.,2013
Revised:
10,Apr., 2013
Accepted:
19,Apr., 2013
Published:
15,May, 2013
Keywords
personalized
marketing,
one-to-one
marketing,
target
marketing,
multi-
campaign,
response
suppression
pattern









Abstract Companies have made their
best efforts to sustain ongoing relations with
their customers and thus generate as much
revenue and profits as possible through
Customer Relationship Management (CRM).
One of the key marketing tools of CRM is
target marketing. Target marketing is a
marketing strategy that analyzes certain
goods and services preferred by existing
customers, develops new goods and services
that fit with the customers' preferences, and
targets those goods and services. As many
companies carry out personalized target
marketing, most customers may encounter
several similar marketing campaigns
simultaneously, thereby making the
campaigns less effective and efficient than
specifically-personalized campaigns.
Furthermore, when a customer is offered
several marketing campaigns, his or her
response is highly dependent on the total
number of campaigns recommended to him
or her. With this being said, this research
shall suggest an optimized strategy to deliver
more effective and efficient multicampaigns
and hence maximize customer satisfaction by
applying different response suppression
functions into various customer groups.


1. Introduction
A dramatic increase in Internet users and an advancement
of IT (Information Technology) have brought us new
theories and concepts in many academic fields. We have
been stressing the significance of word-of-mouth marketing
since the days of old school commerce. Word of mouth
refers to oral communication and the passing of key
information about certain goods such as price, quality and
reliability from satisfied customers to potential customers.
This traditional marketing strategy has been coupled with
dramatic advancements in the Internet and IT and has
created a new marketing terminology, Customer

1,
Gyung-Mok Yoon is an Assistant Professor with Department of
Computer Science and Engineering, Seoul Cyber University, Seoul, Korea
(Email: gihyoon@iscu.ac.kr)
2
Yong-Hyuk KiM is an Associate Professor with Department of
Computer Science and Engineering, Kwangwoon University, Seoul, Korea
(Email: yhdihfly@kw.ac.kr)
3
Byung-Ro Moon is a Professor with School of Computer Science and
Engineering, Seoul National University, Seoul, Korea (Email:
mihoon@isnu.ac.kr)
Relationship Management (CRM)[1]. There has been a
wide spectrum of studies and research on enhancement of
relational marketing using the CRM system. A customer's
purchasing traits and patterns cannot be recorded and kept
in case of general commerce; however, those customer
purchase data can be stored in a database in the case of an
e-commerce. As e-commerce becomes more and more
important, it is essential for companies to utilize the
customer information accumulated in e-commerce and
efficiently manage its customers through CRM. To
maximize their profits and customer satisfaction, most
companies have been implementing one-to-one marketing
or target marketing in order to provide personalized services
tailored to their customers. There has been much research
on prediction of user preference via such one-to-one or
target marketing[2,3,4]. To promote their goods and
services, companies establish and deliver many marketing
campaigns. As a result, most customers are living in a flood
of similar promotion/campaign e-mails. These simultaneous
campaigns are called an overlapped recommendation
problem. The greater the number of recommended
campaigns, the less efficient they are. According to research
by Cannon and Riodan [5] on repeated exposure to ads and
commercials, once the same ads and commercials are
repeated beyond a certain decent level, the effects of the ads
and commercials shall be adversely impacted, and most
customers will become bored and irritated.
Applying the findings that simultaneous recommendation
of several marketing campaigns to one customer adversely
influences the customer's response (overlapped
recommendation problem), Kim and Moon suggested a
multicampaign assignment problems that is associated with
how to optimize marketing recommendation of
multicampaign to multiple customer groups[8-10]. In
addition, they formulated a mathematical model for the
problem and used various heuristic algorithms to solve it. In
this way, they came up with the best assignment method
that applies the heuristic algorithms and improves the
existing methods[8,11]. Kim and Yoon[12] recently
showed this problem is computationally hard, i.e., NP-hard,
so supported the validity of using heuristic search.
As there has recently been soaring interest in
multi-campaigning, we have seen diverse research in many
directions. This recent research includes optimal assignment
of multicampaigns using multi-agent approach[13], swarm
intelligence[14], honey bees[15], differential evolution[16],
and so forth.
In the assignment of multicampaigns with the overlapped
recommendation problem in consideration[10], customer
Multiple Personalized Recommendations with Various
Response Suppression Patterns
Gyung-Mok Yoon
1,

, Yong-Hyuk Kim
2
, & Byung-Ro Moon
3

International Journal of Advanced Computer Science, Vol. 3, No. 6, Pp. 267-273, Jun., 2013.
International Journal Publishers Group (IJPG)


268
preference, from the customer's perspective, diminishes
when there are too many overlapped campaigns. In the
existing research, response suppression functions as per
declining user preference are simply assumed to be equal
for every customer. This research assumes that each
customer group should have its own response suppression
function and calculated different response suppression
functions per customer group. Customers were classified in
groups by the values derived from the k-means clustering
algorithm[6,7]. With an assumption that a customer group
in the same cluster has the same response suppression
function value, there will be k response suppression
functions for k clusters. Customers in the same group have
the same response suppression function. Through these tests,
better results could be attained than in the case where there
is only one response suppression function.
The remainder of this research is comprised of the
following: Section 2 shall touch on the incentives to use
multiple response suppression functions in the assignment
of multicampaigns and their application methods in detail.
Section 3: Algorithms used to resolve the problem; Section
4: Methods used to compute multiple response suppression
functions; Section 5: Results from actual data, and finally
Section 6: Conclusion.

2. Multiple Response Suppression Func-
tions for Assignment of Multicampaigns
This section shall discuss the effectiveness of using
multiple response suppression functions in assigning
multicampaigns and eventually realizing more accurate
multi-campaigning.

A. Effectiveness of the Multiple Response Suppression
Function
If we assume a company recommends three campaigns
to customers and assign each campaign to one customer, we
shall come up with results shown in Tables 1 through 3.
Table 1 depicts the customer preference. User 1 belongs to
the cluster with response suppression function B in Table
3, while User 2 belongs to the cluster with response
suppression function A in Table 3. With these factored in,
Table 2 shows the relevant values if there is only one
response suppression function A in Table 3. Only one
response suppression function in Table 3 means that users
have the same response, while response suppression
functions A and B mean different responses per user.

TABLE 1
CUSTOMER'S CAMPAIGN PREFERENCE

Campaign 1 Campaign 2 Campaign 3
User 1 90 50 85
User 2 60 95 40

TABLE 2
EXAMPLES OF MULTICAMPAIGNS ASSIGNMENT

Campaign 1 Campaign 2 Campaign 3
User 1 1 0 1
User 2 0 1 0
If there is only one response suppression function A in
Table 3, two campaigns are recommended for User 1
whereas one campaign is recommended for User 2.

TABLE 3
CUSTOMER'S RESPONSE SUPPRESSION FUNCTION VALUES

Campaign 1 Campaign 2 Campaign 3
Response A 1.0 0.9 0.7
Response B 1.0 0.4 0.1
In this case, the application of preference and response
suppression values results in 90*0.9 + 85*0.9 + 95*1.0 =
252.5, but this value should be regarded as a biased value. If
we discover that there are two response suppressions, we
then apply the response suppression function values in
Table 3 and re-compute the formula. The result is 90*0.4 +
85*0.4 + 95*1.0 = 165.0. This value is the accurate value
divided by the two clusters.

B. Problem Formalization
From this section going forward, this research shall
prescribe more accurate and detailed multicampaign
assignment problems as per each problem. Forms used by
Kim and Moon[10] will be adopted. Let us assume N to be
the number of customers, { } N B ,..., 2 , 1 = to be a set of
customer groups, { } K C ,..., 2 , 1 = to be a set of cluster
numbers, and M to be the number of campaigns. Inputs,
outputs, constraint conditions and evaluation functions shall
be explained in detail.

1) Input
Each customer shall be given with his or her preference
for each campaign and relevant weighting. The response
suppression function associated with overlapped
recommendation for customers in Cluster C is denoted as
R
c
.
) , 0 [ : ,..., ,
2 1
B f f f
M
: Each campaign's preference
function (actually vector). This research shall take into
consideration only the preferences denoted in positive real
numbers. Preference toward a campaign can be obtained
through on existing preference estimation method such as
CF.
) (i f
j
refers to an estimated preference value of
Customer i on Campaign j.
| | K N C : : Cluster number C(i) means a cluster
number with Customer i.

| | 1 , 0 :
) (
N R
i C
: Response suppression function related
to the number of recommendations. Function value
) (
) (
t R
i C
is defined as a preference diminishing factor when
Customer i is exposed to t recommendations. If
i
H
is
equivalent to the number of recommendations for Customer
i, then actual estimation of preference of Customer i on
Campaign j is equal to
) ( ) (
) (
i f H R
j i i C

.

M
w w w ,..., ,
2 1
: Each campaign's weighting. (Weighting
of Campaign j is a positive value denoted by
j
w
.) Each
campaign j has its own weighting
j
w
. The importance of a
campaign depends on this weighting. In formalized
Yoon et al.: Multiple Personalized Recommendations with Various Response Suppression Patterns.
International Journal Publishers Group (IJPG)


269
multicampaign assignment problems, the preference of
Customer i for Campaign j becomes
) ( ) (
) (
i f H R w
j i i C j

.

2) Constraint Conditions
Constraint conditions and the problem should include the
minimum and the maximum values of the number of
recommendations for each campaign. The number of
recommendations for Campaign j should lie between
j
P

and
j
P . Here,
j
P
and
j
P are the minimum and the
maximum recommendations for Campaign j, respectively.
The difference between these two values,
j
j
P P
, can be
either high or low.

3) Output
Outputs can be denoted by
) (
ij
a A =
, binary campaign
assignment matrix. Here,
ij
a
refers to the assignment of
Campaign j into Customer i. If Campaign j is assigned to
Customer i, then
1 =
ij
a
.

4) Evalation
The sum of preferences in Campaign j is defined as actual
estimation of the recommended customers' preferences in
Campaign j. Optimal level F of campaign assignment
matrix A becomes the sum of weighted preferences in all the
campaigns as follows:
( )

= = e
|
|
.
|

\
|
=
M
j a B i
j i i C j
ij
i f H R w A F
1 1 ;
) (
) ( ) ( ) (


The purpose of the problem is to find matrix A that
maximizes F.

3. Algorithms Employed

A. Heuristic Algorithm
This research shall apply the heuristic algorithm used by
Kim and Moon[10]. This algorithm begins with a
no-recommendation condition and continues to assign
multiple campaigns to customers in a greedy way. It is
called a CAA (Constructive Assignment Algorithm).
Cluster C, Customer i, and Campaign j are used to define
g
C
(i,j) as the quantity of gains optimized through the
assignment of Campaign j to Customer i included in Cluster
C. Initially, gain g
C
(i,j) is equal to w
j
f
j
(i), a multiplication of
campaign weighting and the preference of Customer i for
Campaign j. Generally, gain values are expressed as
follows:

, ) ( )) ( ( ) 1 ( ) , (
) ( ) ( i i i C j j i i i C C
H R i f w H R j i g o o + + =


where
i
o refers to the sum of weighted preferences for the
recommended campaigns by Customer i. This value is
usually a real number, and this research shall use balanced
binary search tree, e.g., AVL tree, for efficient control on
gain values.
We also applied the iterative improvement heuristic
proposed in [10]. After every customer is assigned a proper
number of campaigns, we run this heuristic. It proceeds in a
series of passes. During each pass, the heuristic improves
on an initial assignment to create a new assignment. A pass
of iterative improvement is performed for each campaign.
Given a campaign, the heuristic chooses an equal-sized
subset pair of recommended customers and
non-recommended customers that has the maximum gain
sum when swapped. After swapping the subset pair, another
pass is then executed starting with the new assignment. This
process is repeated until no improvement can be obtained.
By adopting this local improvement heuristic after CAA,
we could obtain 0.03% performance improvement over our
preliminary results of [17].

B. k-means Clustering Algorithm
This research applied a simple k-means clustering
algorithm in order to group the customers with similar
characteristics. Since each customer has his/her own
preference value for M campaigns, he/she can be regarded
as a dot in M-dimensional space. Therefore, there are N dots
spread in M-dimensional space because the number of
customers is N. Clustering these customers into k groups
means classifying N dots into k groups.
The algorithm is used in this research in the following way:
k centers are randomly chosen in M-dimensional space.
Choose a center with the closest distance out of the k
centers. In this way, every customer shall be grouped in at
least one of the k centers. Let us say that a customer group
which selects a center p is
p
C
. Average out the dots in
p
C

per dimension and move the center p into the location of the
averaged dots. Repeat the computation for every p from 1
through K and move each center p into the location of the
computed dots. Then, choose the closest center and repeat
the procedures. At this time, we should terminate the
algorithm if the location of the centers do not change. Using
this method, we can cluster the customers into K groups
which share similar characteristics.

4. Application of Multiple Response
Suppression Functions to
Multi-campaign

A. Test Method
We determine the number of customer clusters. If there
are N customers with M campaigns, then the number of
response suppression functions for the campaigns with K
clusters shall be equal to K.
Filed data obtained from e-mail marketing were used as
test data. Estimated preference values obtained through
collaborative filtering of those field data with 48,559
customers and 33 campaigns were utilized[18,19]. Data sets
used in this research were provided by Optus Inc. Personal
data are comprised of many independent variables and their
International Journal of Advanced Computer Science, Vol. 3, No. 6, Pp. 267-273, Jun., 2013.
International Journal Publishers Group (IJPG)


270
subordinate variables that indicate customer response to
corresponding e-mails. These customers were grouped into
two, three, six, and 10 clusters and tested using the k-means
clustering algorithm. Clustering into 2 groups results in the
two clusters with 33 campaign center values. Likewise,
clustering into three, six and 10 groups shall generate three,
six, and 10 clusters, respectively, with 33 campaign center
values. If we sum and average out all these values, we can
see the different results. Different average values mean that
clustering is successfully accomplished.

B. Multiple Response Suppression Functions
Analysis of real data via clustering showed different
responses per cluster. Cluster center values were used to
compute the response suppression functions per each cluster.
This research sorted 33 center values for each cluster in
descending order, set the first campaign as 1.0, multiplied
the first value appearing in the descending order by 1.0 and
defined the result as the second response value. The
preference computation method shown in the table below
first set the default value as 1.0 and computed the second
response value as (1)*(2)=(4). (3)*(4)=(6) was calculated
for the response value in case of three campaigns. This
computation was repeated until the response suppression
function value reached a value greater than 0.001. In this
way, the response suppression functions for 3, 6 and 10
clusters could be calculated: 1) the number of response
suppression functions for 3 clusters = 3; 2) the number of
response suppression functions for 6 clusters = 6; and 3) the
number of response suppression functions for 10 clusters =
10.

5. Experiment Results

A. Experiment 1: 2 Clusters
This experiment grouped 48,559 customers into two
clusters. In this case, there will be two response suppression
functions formed. However, the experiment assumed that
there was only one response suppression function R
1
and
came up with 30761.62 by using a multi-assignment table
and computing for a preference multiplication response
suppression value. This value is a biased value as explained
in Section 2.1. R
2
value 30209.76 in Table 4 was computed
on the assumption that there was only one response
suppression function R
1
and is therefore also a biased value.
This is the value derived from the following procedures: 1)
we believe there is only one response suppression function
R
1
, 2) a multicampaigns assignment table is then prepared
based on our belief, 3) we come to realize that there are two
response suppression functions, and 4) we finally apply the
two response suppression functions in the
previously-prepared table. This derived value is also a
biased value, but a bit more accurate than the value derived
from the assumption that there is only one response
suppression function. The multicampaigns assignment table
was prepared on the assumption that there are two response
suppression functions with 48,559 customers in two clusters.
Then, 30544.21 resulted from by multiplying the preference
function value by the response suppression function value
as shown in Table 5. This value is a more precise value and
is greater than 29774.27 that was computed on the
assumption that there were two responses up by 769.94.
Each customer group classified via the k-means clustering
algorithm as per customer characteristics has its own
response suppression function. Therefore, more efficient
campaigning can be achieved if different response
suppression functions are applied.

TABLE 4
BIASED RESULTS FOR 2 CLUSTERS
Response
'Only One Response
Suppression Function'
Assumption
'2 Response Suppression
Functions' Assumption
R1 30209.76 30209.76
R2 30765.62 29338.78
Average Biased Value 29774.27

TABLE 5
PRECISE RESULTS FOR 2 CLUSTERS

Response Suppression
Functions Used
Preference * Response
Cluster 1, 2 R
1
, R
2
30544.21


Fig. 1 Response suppression function for 2 clusters

Figure 1 shows a graph of a response suppression
function for two-clusters case. As shown in the figure, the
Response 1 Graph has a gentler slope than that of the
Response 2 Graph. This implies that a customer group in 1
cluster(see Response 1) is better fit for multicampaigns
assignment than is that in 2 cluster(see Response 2).
Experiment results shown below represent the results with
3, 6 and 10 clusters.

B. Experiment 2: 3 Clusters
This experiment grouped 48,559 customers into three
clusters. In this case, there will be three response
suppression functions formed. However, the experiment
assumed that there was only one response function R
1
and
came up with 31412.74 by using a multi-assignment table
and computing for a preference multiplication response
value. This value is a biased value as explained in Section
2.1. R
2
value 30222.23 in Table 6 was computed on the
assumption that there was only one response suppression
function R
1
and is therefore also a biased value. This is the
Yoon et al.: Multiple Personalized Recommendations with Various Response Suppression Patterns.
International Journal Publishers Group (IJPG)


271
value derived from the following procedures: 1) we believe
there is only one response suppression function R
1
, 2) a
multicampaigns assignment table is then prepared based on
our belief, 3) we come to realize that there are three
response suppression functions, and 4) we finally apply the
three response suppression functions in the
previously-prepared table. This derived value is also a
biased value, but a bit more accurate than the value derived
from an assumption that there is only one response
suppression function. The multicampaigns assignment table
was prepared on the assumption that there are three
response suppression functions with 48,559 customers in
three clusters. Then, 30775.44 resulted from multiplying the
preference function value by the response suppression
function value as shown in Table 7.
This value is a more precise value and is greater than
29854.71 that was computed on the assumption that there
were three responses up by 920.73. Each customer group
classified via the k-means clustering algorithm as per
customer characteristics has its own response suppression
function. Therefore, more efficient campaigning can be
achieved if different response suppression functions are
applied.
Figure 2 shows a graph of a response suppression
function for three-clusters case. As shown in the figure, the
Response 1 Graph has a gentler slope than that of the
Response 2 Graph and 3 Graph. This implies that a
customer group in 1 cluster(see Response 1) is better fit for
multicampaigns assignment than is that in 2(see Response 2)
and 3 cluster(see Response 3).

TABLE 6
BIASED RESULTS FOR 3 CLUSTERS
Response
'Only One Response
Suppression Function'
Assumption
'3 Response Suppression
Functions' Assumption
R1 31412.74 29132.15
R2 30222.23 30222.23
R3 30209.76 30209.76
Average Biased Value 29854.71

C. Experiment 3: 6 Clusters
This experiment grouped 48,559 customers into six
clusters. In this case, there will be six response suppression
functions formed. However, the experiment assumed that
there was only one response suppression function R
1
and
came up with 30209.76 by using a multi-assignment table
and computing for a preference multiplication response
value. This value is a biased value as explained in Section
2.1. R
2
value 30229.22 in Table 8 was computed on the
assumption that there was only one response suppression
function R
1
and is therefore also a biased value. This is the
value derived from the following procedures; 1) we believe
there is only one response suppression function R
1
, 2) a
multicampaigns assignment table is then prepared based on
our belief, 3) we come to realize that there are six response
suppression functions, and 4) we finally apply the six
response suppression functions in the previously-prepared
table.

TABLE 7.
PRECISE RESULTS FOR 3 CLUSTERS

Response Suppression
Functions Used
Preference * Response
Cluster 1, 2, 3 R
1
, R
2
, R
3
30775.44


Fig. 2 Response suppression function for 3 clusters

This derived value is also a biased value, but a bit more
accurate than the value derived from then assumption that
there is only one response suppression function. The
multicampaigns assignment table was prepared on the
assumption that there are six response suppression functions
with 48,559 customers in 6 clusters. Then, 30816.11
resulted from multiplying the preference function value by
the response suppression function value as shown in Table 1.
This value is a more precise value and is greater than
30125.17 that was computed on the assumption that there
were six responses up by 690.94. Each customer group
classified via k-means clustering algorithm as per customer
characteristics has its own response suppression function.
Therefore, more efficient campaigning can be achieved if
different response suppression functions are applied.

TABLE 8
BIASED RESULTS FOR 6 CLUSTERS
Response
'Only One Response
Suppression Function'
Assumption
'6 Response Suppression
Functions' Assumption
R1 30209.76 30209.76
R2 30229.22 30229.22
R3 30315.55 30127.88
R4 30228.19 30228.19
R5 30218.71 30218.71
R6 31575.28 29737.27
Average Biased Value 30125.17

TABLE 9
PRECISE RESULTS FOR 6 CLUSTERS

Response Suppression
Functions Used
Preference *
Response
Cluster 1, 2, 3, 4, 5, 6 R
1
, R
2
, R
3
, R
4
, R
5
, R
6
30816.11

Figure 3 shows 2 graphs of a response suppression
function for 6-clusters case. As shown in the figure, the
R
e
s
p


Number of campaigns
Resp
onse
1
Resp
onse
2
International Journal of Advanced Computer Science, Vol. 3, No. 6, Pp. 267-273, Jun., 2013.
International Journal Publishers Group (IJPG)


272
Response 1 and 4 Graph have a gentler slope than that of
the Response 2, 3, 5, and 6 Graph. This implies that a
customer groups in 1(see Response 1) and 5 cluster(see
Response 5) are better fit for multicampaigns assignment
than are those other clusters.

D. Experiment 4: 10 Clusters
This experiment grouped 48,559 customers into 10
clusters. In this case, there will be 10 response suppression
functions formed. However, the experiment assumed that
there was only one response suppression function R
1
and
came up with 30228.19 by using a multi-assignment table
and computing for a preference multiplication response
value. This value is a biased value as explained in Section
2.1. R
2
value 30209.76 in Table 10 was computed on the
assumption that there was only one response suppression
function R
1
and is therefore also a biased value. Moreover,
R
3
, R
4
, R
5
, R
6
, R
7
, R
8
, R
9
and R
10
values are all biased
values.


Fig. 3 Response suppression function for 6 clusters


TABLE 10
BIASED RESULTS FOR 10 CLUSTERS
Response
'Only One Response
Suppression Function'
Assumption
'10 Response Suppression
Functions' Assumption
R1 30228.19 30228.19
R2 30209.76 30209.76
R3 31231.37 30129.44
R4 30229.52 30229.12
R5 33542.61 29127.82
R6 30887.02 30203.61
R7 30222.20 30222.20
R8 30379.85 30184.63
R9 30543.07 30115.52
R10 31116.75 30154.24
Average Biased Value 30080.45

This is the value derived from the following procedures:
1) we believe there is only one response suppression
function R
1
, 2) a multicampaigns assignment table is then
prepared based on our belief, 3) we come to realize that
there are 10 response suppression functions, and 4) we
finally apply the ten response suppression functions in the
previously-prepared table. This derived value is also a
biased value, but a bit more accurate than the value derived
from the assumption that there is only one response
suppression function. The multicampaigns assignment table
was prepared on the assumption that there are 10 response
suppression functions with 48,559 customers in 10 clusters.
Then, 31431.02 resulted from multiplying the preference
function value by the response suppression function value
as shown in Table 11. This value is a more precise value
and is greater than 30080.45 that was computed on the
assumption that there were 10 responses up by 1350.57.
Each customer group classified via k-means clustering
algorithm as per customer characteristics has its own
response suppression function.

TABLE 11
PRECISE RESULTS FOR 10 CLUSTERS

Response Suppression
Functions Used
Preference * Response
Cluster 1, 2, 3, 4, 5,
6, 7, 8, 9, 10
R
1
, R
2
, R
3
, R
4
, R
5
,
R
6
, R
7
, R
8
, R
9
, R
10

31431.02

Therefore, more efficient campaigning can be achieved if
different response suppression functions are applied. Figure
4 shows 2 graphs of the response suppression functions for
ten-clusters case. As shown in the figure, the Response 1
and 6 Graphs have gentler slopes than those of the Response
2, 3, 4, 5, 7, 8, 9 and 10 Graphs. This implies that customer
groups in 1(see Response 1) and 6 cluster (see Response 6)
are better fit for multicampaigns assignment than are those
in other clusters.




Fig. 4 Response suppression function for 10 clusters

R
e
s
p

Number of campaigns
Res
pon
se1
Res
pon
se2
R
e
s
p

Number of campaigns
Res
pon
se6
Yoon et al.: Multiple Personalized Recommendations with Various Response Suppression Patterns.
International Journal Publishers Group (IJPG)


273
6. Conclusion
Based on the fact that repeated ads and commercials
may provoke customer boredom and adversely impact their
effectiveness[5], this research used the k-means clustering
algorithm in order to cluster customers into certain groups
with similar characteristics (i.e., response, preference, etc),
to apply the clustered customer's own response suppression
function and to carry out multicampaigns. Corresponding
results indicated that grouping into 2, 3, 6 and 10 clusters
had 2.59%, 3.08%, 2.29% and 4.49%, respectively, higher
preference * response value than using a single response
suppression function. This implies that better results can be
attained if different response suppression functions are used
for each customers.
Since it is clear that repeated ads and commercials in
one-to-one target marketing adversely influence the
effectiveness and efficiency of the marketing
recommendation, there should be more in-depth research on
customer response in the multicampaigns assignment that
utilizes a campaign-recommending company's preference
and reliability.

Acknowledgments
A preliminary version of this paper appeared in
Proceedings of the International Conference on e-Learning,
e-Business, Enterprise Information Systems, and
e-Government, pp. 405-411, 2011. The ICT at Seoul
National University provides research facilities for this
study. This work was supported by the Engineering
Research Center of Excellence Program (Grant
2011-0000966), Basic Science Research Program (Grant
2012-0001855), and Midcareer Researcher Program (Grant
2011-0018006) of Korea Ministry of Education, Science
and Technology (MEST) / National Research Foundation of
Korea (NRF).

References
[1] J. Dyche, The CRM Handbook: A Business Guide to
Customer Relationship Management, Addison-Wesley,
2001.
[2] G. Adomavicius and A. Tuzhilin, Toward the next
generation of recommender systems: a survey of the
state-of-the-art and possible extensions, (2005) IEEE Trans.
on Knowledge and Data Engineering, vol.17, no.6,
pp.734-749.
[3] J. F. Baldwin, T. P. Martin, and A. Tzanavari, An
intelligent method for inferring information about the
user/client, (2001) Proc. IS, pp.72-79.
[4] J. L. Herlocker, J. A. Konstan, L. G. Terveen, and J. T. Riedl,
Evaluating collaborative filtering recommender systems,
(2004) ACM Trans. on Information Systems, vol.22, no.1,
pp.5-53.
[5] H. M. Cannon and E. A. Riordan, Effective reach &
frequency: Does it really make sense?, (1994) Journal of
Advertising Research, vol.41, pp. 19-28.
[6] J. B. MacQueen, Some methods for classification and
analysis of multivariate observations, In Proceedings of the
5th Berkeley Symposium on Mathematical Statistics and
Probability. University of California Press1967, pages 281
297.
[7] K. Alsabti, S. Ranka, and V. Singh, An efficient k-means
clustering algorithm, (1998) IPPS/SPDP Workshop on High
Performance Data Mining.
[8] Y.-H. Kim and B.-R. Moon, Lagrange multiplier method
for multi-campaign assignment problem, (2004) In
Proceedings of the Genetic and Evolutionary Computation
Conference, volume 2, pages 1065-1077.
[9] Y.-H. Kim and B.-R. Moon, Optimization of multiple
campaigns reflecting multiple recommendation issue, (2005)
Journal of Korea Information Science Society B: Software
and Applications, vol.32, no.5, pp.335-345.
[10] Y.-H. Kim and B.-R. Moon, Multicampaign assignment
problem, (2006) IEEE Trans. on Knowledge and Data
Engineering, vol.18, no.3, pp. 405-414.
[11] Y.-H. Kim, Y. Yoon and B.-R. Moon, A Lagrangian
approach for multiple personalized campaigns, (2008) IEEE
Trans. on Knowledge and Data Engineering, vol.20, no.3,
pp.383-396.
[12] Y.-H. Kim and Y. Yoon, New Theoretical Findings in
Multiple Personalized Recommendations, (2010) In
Proceedings of the Annual ACM Symposium on Applied
Computing, pp. 94-98.
[13] S. Dehuri, S.-B. Cho and A. Ghosh, Wasp: A multi-agent
system for multiple recommendations problem, (2008)
IEEE 4
th
International Conference on Next Generation Web
Services Practices.
[14] S. Dehuri and S.-B. Cho, A novel particle swarm
optimization for multiple campaigns assignment problem,
(2008) In Proceedings of the 5th International Conference
on Soft Computing as Transdisciplinary Science and
Technology.
[15] S. Dehuri, S.-B. Cho and A. K. Jagadev, Honey bee
behavior: A multi-agent approach for multiple campaigns
assignment problem, (2008) IEEE International Conference
on Information Technology.
[16] L. Pradhan, B. B. Mohapatra, S. Dehuri, and A. K. Panda,
The enhancement of the e-readiness ranking of a nation by
solving the multiple campaign assignment problem using
differential evolution, (2009) International Journal of
Electronic Finance, vol.3, no.1, pp.77-95.
[17] G.-M. Yoon, Y.-H. Kim, and B.-R. Moon, Multicampaign
assignment reflecting group-based response suppression,
(2011) In Proceedings of the International Conference on
e-Learning, e-Business, Enterprise Information Systems, and
e-Government, pp. 405-411.
[18] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J.
Riedl, GroupLens: An open architecture for collaborative
filtering of netnews, (1994) In Proceedings of CSCW '94,
Chapel Hill, NC.
[19] U. Shardanand, and P. Maes, Social information filtering:
algorithms for automating 'Word of Mouth', (1995) In
Proceedings of CHI '95. Denver, CO..

You might also like