Customer Clustering Based on Customer Purchasing Sequence Data

Yen-Chung Liu. Int. Journal of Engineering Research and Application www.ijera.com
ISSN : 2248-9622, Vol. 7, Issue 1, ( Part -1) January 2017, pp.49-58
www.ijera.com 49 | P a g e
Customer Clustering Based on Customer Purchasing Sequence
Data
Yen-Chung Liu*, Yen-Liang Chen**
*(Department of Information Management, National Central University, Chung-Li, Taiwan 32001, R.O.C.)
** (Department of Information Management, National Central University, Chung-Li, Taiwan 32001, R.O.C.)
ABSTRACT
Customer clustering has become a priority for enterprises because of the importance of customer relationship
management. Customer clustering can improve understanding of the composition and characteristics of
customers, thereby enabling the creation of appropriate marketing strategies for each customer group.
Previously, different customer clustering approaches have been proposed according to data type, namely
customer profile data, customer value data, customer transaction data, and customer purchasing sequence data.
This paper considers the customer clustering problem in the context of customer purchasing sequence data.
However, two major aspects distinguish this paper from past research: (1) in our model, a customer sequence
contains itemsets, which is a more realistic configuration than previous models, which assume a customer
sequence would merely consist of items; and (2) in our model, a customer may belong to multiple clusters or no
cluster, whereas in existing models a customer is limited to only one cluster. The second difference implies that
each cluster discovered using our model represents a crucial type of customer behavior and that a customer can
exhibit several types of behavior simultaneously. Finally, extensive experiments are conducted through a retail
data set, and the results show that the clusters obtained by our model can provide more accurate descriptions of
customer purchasing behaviors.
Keywords: Customer clustering, Customer segmentation, Sequence data, NMF
I. INTRODUCTION
Customers are the most critical asset of
companies, and customer churn and the subsequent
loss of revenue must therefore be avoided. An
increasing number of companies have invested
resources in customer relations management (CRM)
to quickly adapt their strategies and retain customers.
CRM comprises four elements: customer
identification, customer attraction, customer
retention, and customer development[1]. The first
task among them is customer identification, which
can be further divided into two tasks according to
purpose. The first task is target customer analysis,
which involves finding customers who will be
profitable for a specific company. The second task is
customer segmentation, which involves clustering
customers according to their similar characteristics.
The rapid development of e-commerce has
led to more customer-related data, such as customer
profile, transaction, and Web browsing data, being
collected and stored in databases. By using data-
mining techniques to analyze these data, we can
discern the differences between customers and
identify target customers. Clustering is the most
commonly used data-mining method for customer
identification.
Clustering is a common technique for data
analysis in numerous fields including machine
learning, pattern recognition, information retrieval,
bioinformatics, and various areas of commerce. The
task of clustering is to partition a set of objects into
multiple groups, with the objects in the same group
demonstrating high similarity and the objects in
different groups display in glow similarity. This
technique is usually applied to determine the major
groups of customers in customer segmentation. For
each customer group, appropriate sales plans or
product development strategies can be suggested
according to its characteristics.
In general, customer clustering is executed
according to the different types of customer data
involved, namely customer value data, customer
profile data, customer transaction data, and customer
purchasing sequence data. These data types are
briefly introduced as follows.
Customer value data. In this approach,
customers are clustered according to their value to
the company. First, features are extracted from
customer data, which can help us to determine to
what degree a customer contributes to the company.
A popular model used in this approach determines
customer value according to three extracted features,
namely recency, frequency, and monetary, and is thus
referred to as the RFM model.
Customer profile data. The assumption
behind this approach is that customers with the same
background are more likely to display the same
consumption behaviors than those with different
backgrounds. Accordingly, features derived from
demographic characteristics or background factors
RESEARCH ARTICLE OPEN ACCESS

are selected for clustering. For example, the
clustering attributes could be gender, age, income,
educational background, or the duration of Internet
experience. After clustering, all customers are
partitioned into groups with similar profiles or
backgrounds.
Customer transaction data. The popularity
of electronic transactions enables companies to easily
obtain detailed transaction data for each customer,
containing the items purchased and numbers of units.
By aggregating these transaction data, we can learn
the amount of each itema customer purchased. This
aggregated data focuses on item content without
considering temporal orderrelation among
transactions, and thus represents a static description
of consumer behaviors. The similarity between
customers is decided by the similarity of purchases.
The customers who purchased similar items should
be assigned to the same group. In other words, the
customers in each cluster demonstratesimilar static
consumption behaviors. However, because the
behavior ofa real customer constitutes a temporal
sequence of transactions, the weakness of this
approach is that it ignores the dynamic nature of
customer behaviors. This defect may preventthe
clusters obtained from being able to accurately
describe the dynamic aspect of customer behaviors.
Customer purchasing sequence data. In this
category, the behaviors of a customer are represented
by a sequence of items purchased, where an item
could be a real product or a higher-order conceptual
item. Clustering by item sequence is based
onsimilarity. Two customerswith similar item
sequences should be assigned to the same group. The
advantage of clustering on the basis of item sequence
is that it canaccount for dynamic changes in
sequence(i.e., the dynamic nature of customer
behaviors is included for consideration). However, its
weakness lies in that in reality a customer usually
purchases more than one item in a transaction. The
item sequences fail to represent the real purchasing
behaviors of customers. Item sequence clustering is
thereforeanunsuccessful approach to clustering
customers on the basis of their purchasing behaviors.
In this research, our clustering will be
conductedusing customer sequence data. However,in
our model, a customer sequence is extended from a
sequence of items to a sequence of itemsets. The
reason forthis extension is because a customer may
shop multiple times and buy a set of itemseach time.
By accounting for this phenomenon, a sequence in
our model can properly represent the purchasing
sequence of a customer in the real world.
Additionally, unlike traditional research that
assignseach customer to only one cluster, in our
model we allow each customer to be clustered into
multiple clusters or no cluster. The reason forthis
difference is that in the traditional view,each cluster
represents a group of customers with similar
characteristics or behaviors, whereasin our model,
each cluster represents a major type of purchasing
behavior and each customer can simultaneously
displayseveral types of behaviors. In other words, a
traditional model focuses on showing whose
consumption behaviors are similar, but our model
can reveal typical types of customer behaviors and
shows which of them aredemonstrated by an
individual customer.
In this paper, our goal is touse customer
sequence datato derivea set of clusters, which are
sequences of itemsets. In addition, wealso show
which clusters each customer participates in. The
solution methodology consists of three phases. In the
first phase, the original customer sequence data
aretransformed into a matrix of patterns and
customers, where a pattern is an elementary
purchasing behavior. Then, in the second phase, we
apply the nonnegative matrix factorization (NMF)
method to transform the original matrix into two
matrixes; the first matrix shows the content of each
typical purchasing type and the second matrix shows
the relationship between typical purchasing types and
customers. Finally, the third phase assigns each
customer into clusters according to the matrix of
typical purchasing types and customers.
The remainder of this paper is organized as
follows. In Section 2, the background and related
works are reviewed. Our solution methodology is
proposed in Section 3. Section 4 discusses how the
experimentsare performed to validate the results.
Finally, we providea conclusion in Section 5.
II. RELATED WORK
In Section 2.1, we review how previous
research has applied clustering techniques to
customer data. We then introduce the background of
the NMF method in Section 2.2.
2.1.Customer clustering
As mentioned in Section 1, customer
clustering is conducted according to different types
of customer data, including customer value data,
customer profile data, customer transaction data, and
customer purchasing sequence data. In the following
sections, we briefly introduce each type.
2.1.1 Customer value data
Most companies possess some
VIPcustomers who contribute ahigher or more stable
profitto the company than average customers do.
Companies naturally want to identify those VIP
customers and providethem with better care to avoid
churn. Accordingly, a reliable methodof assessing
customer value is essential. The RFM and customer
lifetime value (LTV) models are most popularly used
for customer value evaluation. Through these

models, companies can learn which features should
be extracted from the personal or transaction data of
their customers and use them to assess customer
value . It used three metrics, including the present
value, the potential value, and customer loyalty, to
evaluate customer value by using the LTV theory. A
framework was designed to illustrate how these three
metrics should be calculated. Finally, according to
the values of these three metrics, they clustered
customers into different segments[2],[3],Another
study clustered credit card customers into different
segments by using SOM technology based on
previous repayment behavior and RFM behavioral
scoring predicators[4],Other researcherscompared the
advantages and disadvantages of the customer
clustering results derived withthree approaches,
RFM,chi square automatic interaction detection , and
Logistic regression, for two different client groups,
members of an e-mail-order retailerand a nonprofit
organization. This approachis usually applied to
target customer analysis, with the goalof finding
customers who possess certain characteristics[5].
2.1.2 Customer profile data
In this line of research, the features used are
usually collected from data concerning the
backgrounds or demographic characteristics of
customers. One researcher clustered sightseeing
visitors to Cape Town based on features selected
from tour characteristics, opinions about Cape Town,
and demographic characteristics[6], another
researcherevaluated 40 potential customers of a drink
company on the basis oftheir degree of socialization,
leisure, achievement and other indicators, and
appliedthree different data mining methods to
compare the clustering results, including a support
vector machine, self-organizing feature map
(SOFM), and k-means method[7], the other
researcher proposed a two-stage customer clustering
methodology combined with SOFM and k-means
based on features including user experience, services
provided in an electronic goods store, and
demographic characteristics[8], and other researcher
proposed an integrated method of profitable customer
segmentation for a motor company according to
features collected from demographic, customer
satisfaction, and accounting data[9],The other
researcher developed a soft clustering method that
uses a latent mixed-class membership clustering
approach to classify online customers according to
four types of features, namely consumer network
usage, satisfaction with service, buying behaviors,
and demographic characteristics; notably, because
the clustering resultsare soft rather than hard, a
customer may belong to multiple groups[10].
This line of study is helpful for
understanding into what groups a customer base can
be divided, thereby providing support for product
planning and market promotion. However, because
the characteristics used for clusters may be changed
depending on the application, this approach should
be custom designed according to the specific
demands of the application.
2.1.3 Customer transaction data
Earlier studies have performed customer
segmentation on the basis of demographic
characteristics or background information. However,
clustering customers according to profile data is most
likely no longer useful in the current, increasingly
complicated world. In recent years, the use of
customer transaction or behavioral data for customer
clustering has become a popularresearch subject.
Trading can be conveniently performed through the
Internet, a process that enablesthe easy recording of
customer transactions. Because these transaction data
can reveal customer preferences and needs, research
on the use of transaction data in customer
segmentation is gradually becoming a trend.
some researchers developed a market
segmentation methodology based on product-specific
variables such as items purchased and the associated
monetary expenses from the transactional histories of
customers[11],other researchers represented each
transaction as a vector, with each vector element
corresponding to an item and 0 and 1 representing
the absence and presence of the item in a transaction,
respectively ; they then converted all vectors into an
association matrix, in which they found similarities
between clients and obtained clustering results[12].
The other researchers represented the purchasing
behaviors of each customer through a vector of items
and determined the distance between two items
through the product hierarchy; after the distance
between two customers was computed by
determining the distance between the corresponding
vectors, customers were clustered according to these
distances[13],the other researchers represented
customer data by using a vector of items with
quantity values, determined the similarity of two
customers by the amount of items, and clustered
customers according to their similarities[14].
These studies have determined clusters on
the basis of similarities between customers, whom
were decided according to the similar characteristics
and differences in quantity of the items they
purchased. They ignored the fact that the purchasing
behaviors of a customer are constituted of a sequence
of purchasing decisions made over time. Numerous
patterns can be identified from the entire purchasing
history of a customer, information crucial for
describing and discriminating customer behaviors. If
this dynamic aspect is not considered, the clusters
obtained may be unreliable in practical situations.

2.1.4 Customer sequence data
The advantage of this approach is that the
dynamic aspect of customer behaviors is included by
formattingthe data as a sequence of items purchased
over time. Accordingly, the clusters obtained inthis
model are more reasonable than those obtained
usingprevious static models. Several studies
haveused this model[15],[16].In general, item
sequences may be appliedin areas as varied as
genomic DNA, protein sequences, Web browsing
records, and transaction records.
Some researcher proposed a new measure to
compute the similarity between two sequences and
developed a hierarchical clustering algorithm to
cluster sequence datasets such as protein sequences,
retail transactions, and web-logs[15],other researcher
proposeda method combining model-based and
heuristic clustering for categorical sequences. In the
first step, the categorical sequences are transformed
throughan extension of the hidden Markov model
into a probabilistic space, where a symmetric
Kullback–Leibler distance can operate. In the second
step, the sequences are clustered using hierarchical
clustering on the matrix of distances[16].
In the study of customer sequence data
clustering, an item could be a real, specific product at
the lowestlevel or a conceptual item at a higher level.
Although item sequence clustering seems applicable
to customer purchasing data, the behaviors of a
customer areconstituted by a sequence of itemsets
rather a sequence of items. In a real user transaction,
a user may buy several items at a time;it is unrealistic
to assume that only a single item is purchased in each
transaction. For this reason, the item sequence
clustering approach cannot be applied to real
customer purchasing data.
Traditional research assigns each customer
to only one cluster: all customers are partitioned into
a complete set of nonoverlapping groups, where each
group possesses the same behaviors. By contrast, in
our model, we derive a certain set of clusters from
the behavioral data of all customers. Each cluster
denotes a major type of purchasing behavioror
lifestyle. We accordinglyallow each customer to be
included in multiple clusters or no cluster, because
people can possess diverse lifestyles. For example, a
student with a photography hobby may possess both
the lifestyle of a student and that of a photographer.
In other words, a traditional model focuses on
partitioning customers into groups with similar
behaviors, but our model focuses on revealing typical
types of customer behaviors.
2.2 NMF
Each element in the shopping sequence
usually contains more than one item; therefore, a
traditional sequence clustering method is not
applicable to customer purchasing sequence
clustering. To remedy this problem, the original
customer sequence data aretransformed into a matrix
of elementary patterns and customers, where
anelementary pattern corresponds to an elementary
purchasing behavior. Essentially, an elementary
pattern is a binary tuple, such as (a, b), which means
item b will be boughtdirectly after the purchase of
item a. We then apply the NMF method to transform
this matrix into two matrixes; the first matrix shows
the content of each typical purchasing type and the
second matrix shows the relationship between typical
purchasing types and customers.Specifically, the first
matrix shows how each typical purchasing type is
formed by elementary patterns, whereas the second
matrix shows how a customer consumption behavior
is constituted by typical purchasing types.The NMF
method can help us to extract typical purchasing
types from customer purchasing sequence data.
NMF is a method of multivariate analysis
and linear algebra that is commonly used in latent
semantic analysis. A nonnegative matrix V can be
decomposed into two matrixes, W and H, with V =
WH, where all the values of the elements in the three
matrixes are not negative. Here, matrix V is of the
size m × n, matrix W is of the size m × r, andmatrix
H is of size r × n. The value of r is set by the user,
and is typically r«m and n. A small value of r can
greatly reduce the matrix sizes, which facilitates
further analysis.
Some researcherused the NMF method for
document clustering. They first found all keywords
from documents and then built a matrix V of these
keywords and documents. By using NMF, they
decomposed matrix V into W and H. In their study,
W wasthe matrix of keywords and themes. Similarly,
H wasa matrix of themes and documents, which
demonstrateshow a document is formed by themes.
They assigned a document to the theme with the
largest value in the row of this particular document in
H. Theyalso indicatedthat NMF was more suitable
forfindingthe semantic characteristics of themes than
other methods such as singular value decomposition,
because NMF did not require that the themes be
orthogonal toeach other. Without this restriction,
NMF can focus on finding the themes with the best
clustering effect [17].
Other researcherapplied the NMF method
todocument summarization. They first built a matrix
V of keywords and sentences, and then decomposed
it into two matrixes W and H, where H was a matrix
of topics and sentences. Finally, they chose the
sentences with the highest score for each topic as the
summary of the document.[18]. Because customers
may have diverse lifestyles, clusters canoverlap with
each other and a customer may belong to multiple
clusters. NMF is suitable for solving our situation
because it can indicatemultiple typical purchasing

types and assign a customer to several types
simultaneously.
III. METHODOLOGY
The purchasing sequence of a customer
reveals their preferences or behaviors. When two
customers have different preferences, the
characteristics of their purchasing sequences will
likelydiffer. The types and quantities of items a
customer buys arefrequently used to represent the
characteristics of theirpurchasing behaviors. In
addition tothis static nature of purchasing behaviors,
we can consider how to draw the dynamic features of
customers to describe their behaviors. How theitems
purchased change over time is a key dynamic aspect
that can be discovered from purchasing sequence
data. If this dynamic feature is also used to
discriminate customer behaviors, we can expect a
more reliable clustering result from a more complete
consideration. Accordingly, we analyze what patterns
are frequently repeated in customer purchasing
sequences. These repeated patterns arecrucial for
discerning different user behaviors. When users have
different preferences, different repeated patterns
occur in their respective purchasing sequences. In
other words, when two sequences differ, they must
have different repeated patterns, and vice versa. We
can thus use the repeated patterns in sequences to
discriminate customer behaviors.
In this paper, we adopt elementary patterns
to describe and discriminate customer behaviors.An
elementary pattern is a binary tuple, such as(a, b),
which means that item b will be bought directly after
the purchase of item a. We use elementary patterns to
describe customer behaviors for several reasons.
First, they can be easily collected. Second,
becausetheir form is simple, their usecan reduce the
number of combinations that we must consider,
which in turn reduces the computation cost. Third,
they contain enough information to represent user
behaviors, including the information about what
items we buy and in what quantity and order.
Our algorithm contains three phases. First,
the original customer sequence data aretransformed
into a matrix of elementary patterns and customers,
where an elementary pattern is a repeated unit
purchasing behavior in customer transaction data.
The second phase applies the NMF method to
transform the above matrix into two matrixes, where
the first matrix shows the content of each typical
purchasing type and the second matrix shows the
relationship between typical purchasing types and
customers.The third and final phase assigns each
customer to clusters according to the matrix of
typical purchasing types and customers. In the
following subsections, we describe each phase in
detail.
3.1 Data transformation
The goal of this phase is to build a matrix of
elementary patterns and customers from customer
purchasing sequence data. We first providethe
following definitions.
Definition 1. Let denote the set of
all items occurring in customer sequence data. The j-
th transaction of customer can then be denoted as
， |=s，
| .
Definition 2. Let and be two consecutive
transactions of customer in sequence, and let
denote an item in , and denote an item in
. Then, forms anelementary pattern for
and . Let denote
the set of all elementary patterns generated from
and . Then, .
Definition 3. If customer has n transaction in his
purchasing sequence, all the elementary patterns of
customer can be denoted by
Example. Assume that customer has five
transactions in sequence as follows. A total of 16
elementary patternsare generated as a result.
Table 1. Generating elementary patterns throughtransaction data
The items purchased at the
j-th transaction
Elementary patterns generated
=
=
=
=
=
However, because occurs twice, a total of 15 distinct elementary patterns are generated.In other words,
we have
.
Definition 4. Let the set of customers be denoted as . The elementary patterns of all customers

can then be denoted by , where }.
Definition 5. Let PM denote the matrix of elementary patterns and customers. Specifically, is the number
of times that pattern x appears in the sequence data of customer , where ， .
Table 2. Matrix of elementary patterns between customers
Pattern/Customer
3
2
Definition 6. Let be the normalized matrix of PM. Specifically,
[x,y]= , .
We must normalize the matrix because
customers who purchase large amounts of items tend
to display high frequencies in all elementary
patterns;the same is truefor customers who buy
comparatively less. By normalizing these values with
respect to the total frequency of acustomer, we can
reveal the relative frequencies of patternsofthis
individual customer, which represent their true
preferences.
Table 3.Normalized matrix of PM
Pattern/Customer
0.6
0.4
We want to assign customers to clusters through analyzing matrix . If the behavior
of customer is similar to several clusters, such as , we can make .
3.2 Matrix decomposition through NMF
In the first phase, we retrieved the
elementary patterns from customer purchasing
sequence data, hoping to assess the similarity of the
customers through determining the differences
between the elementary patterns. Because the
appears 3 times in transaction recordappears twice in transaction record
The proportion of in transaction record is 0.6
The proportion of in ransaction record is 0.4

elementary patterns are numerous and
interdependent, they cannot feasibly be used
todirectly evaluate the differences between
customers. Instead, we apply the NMF method to
transform the original matrix V into two
matrixes, where the first matrix W shows the content
of each typical purchasing type and the second
matrix H shows the relationship between typical
purchasing types and customers.
After applying NMF to the matrix PM’ (V),
we obtaintwo matrixes, W and H, by the relation that
= . Here, r is the number
of typical purchasing types specified by a user.
Usually, we set r as equal to the number of groups
asked by user. Matrix W enables us to understand
how each typical purchasing type (group
characteristic[GC]) is formed by elementary patterns.
Similarly, matrix H shows us how the consumption
behaviors of each customer areconstituted by typical
purchasing type (GC). In the following figure, we set
r=2. Matrix V can then be transformed into two
matrixes.
Fig. 1.Relationships amongmatrixes V, W, and H
From matrix H, we learn how the four
customers are related to two different group
characteristics (typical purchasing types).
Additionally, matrix W tells us how each group
characteristic is constituted by elementary patterns.
3.3 Clustering
Table 4.Example of how to generate a clustering threshold
Group
Customer
Customer’s
GC value >0
Sum of
GC value
Average
0 0 0 0.00136 0.02287 3873 29.15666 0.00752
0.00494 0.00755 0.00456 0.01724 0.01857 5652 59.5683 0.01053
0.00011 0.00070 0 0.00047 0 1651 3.96295 0.00240
0.00012 0.00001 0.00009 0 0 2611 5.34028 0.00204
0.00107 0 0.00258 0.00603 0.01154 3984 21.50367 0.00539
Matrix H shown in Fig. 1, shows us thatthe
consumption behaviors of customer areformed by
weight 0.02287 on group characteristics, weight
0.01857 on group characteristics, weight 0.01154
on group characteristics, and zeros on others. How
can we assign customer to groups? Intuitively, if a
customer has a comparatively larger weight in a
certain group, we should assign the customer to this
particular group. Therefore, among all customers
with positive weights in a group, we assign the
customers whose weights are in the p-percentile into
this cluster, where p is set according to the user.
On the basis of this clustering strategy, a
customer is clustered into multiple groups or no
group. Notably, we use the p-percentile as the
threshold; thus, every group is nonempty because
there must besome customers with weights in the p-
percentile.
IV. EXPERIMENTS
To evaluate the performance of the
algorithm, we mustdefine a metric to measure the
qualityof aclustering result. A group of customers
containsa total of customers and elementary
patterns. Let denote the most frequent N
elementary patterns in group . Our hypothesisis
that fora group witha more favorableclustering effect,
more customers shoulddisplaysimilar behaviors. In
other words, there should be more customers in the
group possessing top N frequent patterns. We use
= to
represent the average proportion of customers in
group ,possessing patterns . Additionally,
to measure the interdistance between groups, we
define
to measure the difference between group and
other groups. For example, assume that the number
of clusters is four, and among the top three patterns,
there are twocommon patterns between and ,
onecommon pattern between and , and two

common patterns between and . We therefore
derive =0.4444. An
increasing value of PDP indicates fewer common
patterns between different groups. In other words,
intergroup differences become larger.
To integrate these two indexes, we use the
following index to measure the clustering effect of
group : . Alarger index is
considered more favorable. Finally,we use
CI= to measure the total clustering effect,
which is the weighted average of in all groups.
When the value of CI is large,the purchasing patterns
of the customers in the same group are more similar
with one another and the purchasing patterns of the
customers in different groups are more dissimilar
with one another. When the value is 1,all customers
in eachgroup displaythe same top N patterns, and the
top N patterns of all groups are different.
The data set used in the experiment is public
customer purchasing sequence data collected from a
storeon2001/12/27 and 2002/12/31.The data set
containedinformation for 7,824customers
[19].Among them, 2,096 customers wereone-time
buyers;thus, we remove them from consideration.A
total of 5,728 customers and 54,537 transactionsare
eventually includedin the experiment. The items used
in the experiment are higher-level conceptual items,
to avoid an exceptionally highnumber of patterns
with low support. As a result, we obtain110 items and
generate 10,401 elementary patterns.We use the
software MATLAB toolbox to perform NMF
algorithm. This toolbox was released by Source Code
for Biology and Medicine[20].
The contrast algorithm used in the
experiment is thek-means algorithm, in whicheach
customer is viewed as a data point. Because we have
10,441 elementary patterns, each data point has
10,441 dimensions. However, when the number of
dimensions is excessivelylarge, adimensionality
problem will occur and render the running time
intractable. We therefore choose Q most frequent
elementary patterns in place of the original 10,441
dimensions. To find the most appropriate value of Q,
we test different potential values, including 5, 10, 15,
20, 30, 35, 40, 50, 100, and 200. The results indicate
that when the value of Q is set to 30 or 40, we have
the optimalclustering effect. In this experiment, we
therefore decide to set Q to30.
We additionally set k as5,10,15,20,25, and
30, which are the numbers of clusters used in the
experiment. We use a k-mean algorithm to cluster the
5,728 data points with 30 dimensions into k clusters.
The distance between data points is computed by
using a formula involving the square root of each
element of the array. We use the Pythonprogramming
language to perform a k-means algorithm.
In the experiments, we set the thresholds of
NMF to 50, 60, and 70. In other words, among all
customers with positive weights in a group, we
assign the customers whose weights are in the 50th,
60th, and 70th percentiles to this cluster. In addition,
we set the number of groups as 5, 10, 15, 20, 25, and
30 for NMF and k-means. Furthermore, for the top N
frequent patterns, we set N as 3 and 6. The two
algorithms are executed 30 times for each
combination of parameters. The results for when N
=3 and N = 6 are presented in Tables 5 and 6,
respectively.
Table 5. CI of TOP 3 patterns throughk-means and
the NMF method
K K
Means
NMF
thr=70
NMF
thr=60
NMF
thr=50
5 0.228 0.225 0.238 0.256
10 0.282 0.283 0.297 0.310
15 0.314 0.310 0.327 0.336
20 0.329 0.328 0.344 0.355
25 0.340 0.340 0.357 0.366
30 0.353 0.350 0.364 0.373
The results in Table 5 indicate that when
N=3, NMF has similar performance as that of k-
means if the threshold is 70. However, if the
threshold is 60 or 50, NMF has a more
favorableperformance than k-means does. In
addition, the results show that when the threshold is
lower, the clustering effect of NMF improves,
because when the threshold is lower, we use a higher
standard to assign a customer to a group. This causes
the customers in the same group to exhibitmore
similar behaviors.
The results in Table 6 indicate that when
N=6, NMF has amore favorableperformance than k-
means in all different thresholds. In addition, the
results show that when the threshold is lower, the
clustering effect of NMF improves.
Table 6. CI of TOP 6 patterns throughk-means and
the NMF method
K K
Means
NMF
thr=70
NMF
thr=60
NMF
thr=50
5 0.175 0.193 0.215 0.229
10 0.21 0.228 0.245 0.256
15 0.229 0.244 0.256 0.268
20 0.239 0.254 0.267 0.277
25 0.248 0.263 0.274 0.284
30 0.256 0.268 0.278 0.288

Fig. 2. Change inCI from k=5 to 30 in k-means and NMFwith thr=60%
The results displayed in Fig. 2 compare the
performance ofk-means and NMF with the threshold
of 60. In this comparison, N is set to 3 and 6 and the
number of groups is set to 10, 15, 20, 25, and 30. The
results indicate that when N is smaller, the clustering
performance is more favorable, because when N is
small, only the few most frequent patterns are used
for computing group cohesion.
V. CONCLUSION
In the past, the behaviors of a customer
wereusually represented by a sequence of items
purchased. Clustering by item sequence is instead
based on the similarity of item sequences. The
advantage of clustering based on item sequence is
that the dynamic nature of customer behaviors is
included for consideration. However, its weakness is
that, in reality, a customer usually purchases more
than one item in a transaction. The item sequences
fail to represent the real purchasing behaviors of
customers. Therefore, this research studied the
customer clustering problem on the basis of
transaction sequence rather than item sequence.
Unlike traditional research that assignseach
customer to only one cluster, our model assigns each
customer to multiple clusters or no cluster. The
reason forthis difference is that in the traditional
view, each cluster represents a group of customers
with similar characteristics;by contrast,in our model,
each cluster represents a typical type of purchasing
behavior and each customer can manifest several
types of behavior simultaneously. We thus apply a
different model, NMF, to cluster the transaction
sequence data of all customers. The results obtained
include two matrixes, where the first matrix shows
the typical types of consumption behaviors and the
second matrix shows which of these types a customer
exhibits.
In the experiments, we compare our model
with a traditional k-means algorithm. The results
indicate that the clusters obtained by our model have
higher CI values than do those obtained by k-means.
In other words, compared with the clusters obtained
by k-means, in our model the purchasing patterns of
the customers in the same group are more similar and
the purchasing patterns of the customers in different
groups are more dissimilar.
In the future, this research can be extended
in the following areas. First, in this work,the NMF
model is used for clustering customer transaction
sequence data. However, NMF is a general model for
matrix decomposition and it may not be the most
fitting for the intended work. In the future, we may
attempt to custom design other models thatare more
appropriate for this situation. Second, in this work,
we extract elementary patterns to represent the
original transaction sequence data.
Becauseelementary patterns are onlyone type of
information that can beextracted from sequence data,
we may attempt other methods to extract important
information from the original sequence. We believe
there is a chance that other extracted information can
provide the same or even an increased ability to build
the model. Finally, typical types of customer
behaviors are first introduced in our clustering work.
The problem of how to find these types is more
interesting than the clustering itself. In the future, we
will propose a study that focuses entirely on finding
typical types of customer behaviors in customer
transaction sequence data.

REFERENCES
[1].Ngai, E. W., Xiu, L., & Chau, D. C,
Application of data mining techniques in
customer relationship management: A
literature review and classification, Expert
systems with applications, 36(2), 2009,
2592-2602.
[2].Hwang, H., Jung, T., & Suh, E, An LTV model
and customer segmentation based on
customer value: a case study on the wireless
telecommunication industry, Expert systems
with applications, 26(2), 2004, 181-188.
[3].Kim, S. Y., Jung, T. S., Suh, E. H., & Hwang,
H. S, Customer segmentation and strategy
development based on customer lifetime
value: A case study, Expert systems with
applications, 31(1), 2006, 101-107.
[4].Hsieh, N. C, An integrated data mining and
behavioral scoring model for analyzing
bank customers, Expert systems with
applications, 27(4), 2004, 623-633.
[5].McCarty, J. A., & Hastak, M, Segmentation
approaches in data-mining: A comparison of
RFM, CHAID, and logistic
regression, Journal of business
research, 60(6), 2007, 656-662.
[6].Bloom, J. Z, Market segmentation: A neural
network application,Annals of Tourism
Research, 32(1),2005, 93-111
[7].Huang, J. J., Tzeng, G. H., & Ong, C. S,
Marketing segmentation using support
vector clustering, Expert systems with
applications, 32(2), 2007, 313-317.
[8].Kuo, R. J., Ho, L. M., & Hu, C. M. Integration
of self-organizing feature map and K-means
algorithm for market
segmentation, Computers & Operations
Research, 29(11), 2002, 1475-1493.
[9].Lee, J. H., & Park, S. C, Intelligent profitable
customers segmentation system based on
business intelligence tools, Expert systems
with applications, 29(1), 2005, 145-152
[10]. Wu, R. S., & Chou, P. H, Customer
segmentation of multiple category data in e-
commerce using a soft-clustering
approach, Electronic Commerce Research
and Applications, 10(3), 2011, 331-341.
[11]. Tsai, C. Y., & Chiu, C. C, A purchase-based
market segmentation methodology, Expert
Systems with Applications, 27(2), 2004,
265-276.
[12]. Lu, T. C., & Wu, K. Y, A transaction
pattern analysis system based on neural
network, Expert Systems with
Applications, 36(3), 2009, 6091-6099.
[13]. Hsu, F. M., Lu, L. P., & Lin, C. M,
Segmenting customers by transaction data
with concept hierarchy, Expert Systems with
Applications, 39(6), 2012, 6221-6228.
[14]. Lu, K., & Furukawa, T, A Framework for
Segmenting Customers Based on
Probability Density of Transaction Data,
In Advanced Applied Informatics (IIAIAAI),
2012 IIAI International Conference
on IEEE, 2012, September, 273-278.
[15]. Oh, S. J., & Kim, J. Y, A hierarchical
clustering algorithm for categorical
sequence data, Information Processing
Letters, 91(3), 2004, 135-140.
[16]. De Angelis, L., & Dias, J. G, Mining
categorical sequences from data using a
hybrid clustering method, European Journal
of Operational Research, 234(3), 2014, 720-
730.
[17]. Xu, W., Liu, X., & Gong, Y,. Document
clustering based on non-negative matrix
factorization, In Proceedings of the 26th
annual international ACM SIGIR
conference on Research and development in
information retrieval, 2003, 267-273
[18]. Lee, J. H., Park, S., Ahn, C. M., & Kim, D,
Automatic generic document summarization
based on non-negative matrix factorization,
Information Processing &
Management, 45(1), 2009, 20-34.
[19]. Chen, Y. L., Kuo, M. H., Wu, S. Y., &
Tang, K, Discovering recency, frequency,
and monetary (RFM) sequential patterns
from customers’ purchasing data,Electronic
Commerce Research and Applications,
8(5),2009, 241-251.
[20]. Li, Y., & Ngom, A, The non-negative
matrix factorization toolbox for biological
data mining,Source code for biology and
medicine, 8(1), 2013, 1

Customer Clustering Based on Customer Purchasing Sequence Data

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Customer Clustering Based on Customer Purchasing Sequence Data (20)

Recently uploaded (20)

Customer Clustering Based on Customer Purchasing Sequence Data