6

Data Mining Using RFM Analysis

Derya Birant
Dokuz Eylul University
Turkey

1. Introduction
RFM stands for Recency, Frequency and Monetary value. RFM analysis is a marketing
technique used for analyzing customer behavior such as how recently a customer has
purchased (recency), how often the customer purchases (frequency), and how much the
customer spends (monetary). It is a useful method to improve customer segmentation by
dividing customers into various groups for future personalization services and to identify
customers who are more likely to respond to promotions.
In recent years, data mining applications based on RFM concepts have also been proposed
for different areas such as for the computer security (Kim et al., 2010), for automobile
industry (Chan, 2008) and for the electronics industry (Chiu et al., 2009). Research cases of
data mining with RFM variables include different data mining techniques such as neural
network and decision tree (Olson et al., 2009), rough set theory (Cheng & Chen, 2009), self
organizing map (Li et al., 2008), CHAID (McCarty and Hastak, 2007), genetic algorithm
(Chan, 2008) and sequential pattern mining (Chen et al., 2009; Liu et al., 2009).
Integration of RFM analysis and data mining techniques provides useful information for
current and new customers. Clustering based on RFM attributes provides more behavioral
knowledge of customers actual marketing levels than other cluster analyses. Classification
rules discovered from customer demographic variables and RFM variables provides useful
knowledge for managers to predict future customer behavior such as how recently the
customer will probably purchase, how often the customer will purchase, and what will the
value of his/her purchases. Association rule mining based on RFM measures analyzes the
relationships of product properties and customers contributions / loyalties to provide a
better recommendation to satisfy customers needs.
This chapter presents incorporating RFM analysis into data mining techniques to provide
market intelligence. It proposes a new three-step approach which uses RFM analysis in data
mining tasks, including clustering, classification and association rule mining, to provide
market intelligence and to assist market managers in developing better marketing strategies.
In our model, (i) once clustering task is used to find customer segments with similar RFM
values, (ii) then, using customer segments and customer demographic variables,
classification rules are discovered to predict future customer behaviors, (iii) finally;
association rule mining is carried out for product recommendation. The proposed model
depends on the sentence "the best predictor of future customer behavior is past customer
behavior". (Swearingen, 2009)
The purpose of this study is to provide better product recommendations than simple
recommendations, by considering several parameters together: customers segment, the

www.intechopen.com
92 Knowledge-Oriented Applications in Data Mining

current RFM values of the customer, potential future customer behavior and products
frequently purchased together. To the best of our knowledge, this chapter is the first in
applying the RFM criterion in three data mining tasks, applied one after another, using
customer demographic data, customer transaction data, and product properties.
Experiments, which were carried out using the datasets collected by a sports store in Turkey
through its e-commerce website, empirically demonstrate the benefits of using our model in
direct marketing.
The rest of the chapter is organized as follows. Section 2 introduces the basics of RFM
analysis and explains the recency, frequency and monetary concepts in detail. Section 3
reviews the literature and describes how data mining and RFM analysis are combined in the
previous studies. Section 4 presents our proposed model and describes its architecture in
detail. Section 5 demonstrates how the proposed model can be used to analyze a real world
data, as a case study, including data preprocessing, RFM analysis, customer segmentation,
customer behavior prediction and product recommendation. Finally, Section 6 concludes the
chapter.

2. RFM analysis
The concept of RFM was introduced by Bult and Wansbeek (1995) and has proven very
effective (Blattberg et al., 2008) when applied to marketing databases. RFM analysis depends
on Recency (R), Frequency (F), and Monetary (M) measures which are three important
purchase-related variables that influence the future purchase possibilities of the customers.
Recency refers to the interval between the time, that the latest consuming behavior happens,
and present. Many direct marketers believe that most-recent purchasers are more likely to
purchase again than less-recent purchasers. Frequency is the number of transactions that a
customer has made within a certain period. This measure is used based on the assumption
that customers with more purchases are more likely to buy products than customers with
fewer purchases. Monetary refers to the cumulative total of money spent by a particular
customer.
In order to demonstrate RFM analysis, an example dataset (customer transaction data) is
given in Table 1. Table 2 shows the steps of RFM analysis, which involves scaling customers
based on each RFM factor separately. The segmentation starts with recency, then frequency,
and finally monetary value. It begins with sorting customers based on recency, i.e. period
since last purchase, in order of lowest to highest (most recent purchasers at the top). The
customers are then split into quintiles (five equal groups), and given the top 20% a recency
score of 5, the next 20% a score of 4 and so on. Customers are then sorted and scored for
frequency from the most to least frequent, coding the top 20% as 5, and the less frequent
quintiles as 4, 3, 2, and 1. This process is then undertaken for monetary as well. Finally, all
customers are ranked by concatenating R, F, and M values. This example shows that RFM
analysis can be useful even if database is small of only 15 transactions whereas it would be
more powerful when the database grows.
RFM analysis assigns value-scores to each customer on the basis of her past behavior. Using
the quintile system explained above, at the most, 125 different scores (5x5x5) can be
assigned. These cells differ in size from one another. A customers score can range from 555
being the highest, to 111 being the lowest. The best customers are in quintile 5 for each
factor (555) that have purchased most recently, most frequently and have spent the most
money.

www.intechopen.com
Data Mining Using RFM Analysis 93

CustomerID Recency (Day) Frequency (Number) Monetary (TL)

1 3 6 540
2 6 10 940
3 45 1 30
4 21 2 64
5 14 4 169
6 32 2 55
7 5 3 130
8 50 1 950
9 33 15 2430
10 10 5 190
11 5 8 840
12 1 9 1410
13 24 3 54
14 17 2 44
15 4 1 32
Table 1. An example dataset: customer transactions

CID Rec. R CID Freq. F CID Mon. M CID RFM

12 1 5 9 15 5 9 2430 5 1 544
1 3 5 2 10 5 12 1410 5 2 454
15 4 5 12 9 5 8 950 5 3 111
7 5 4 11 8 4 2 940 4 4 222
11 5 4 1 6 4 11 840 4 5 333
2 6 4 10 5 4 1 540 4 6 222
10 10 3 5 4 3 10 190 3 7 433
5 14 3 7 3 3 5 169 3 8 115
14 17 3 13 3 3 7 130 3 9 155
4 21 2 14 2 2 4 64 2 10 343
13 24 2 4 2 2 6 55 2 11 444
6 32 2 6 2 2 13 54 2 12 555
9 33 1 15 1 1 14 44 1 13 232
3 45 1 3 1 1 15 32 1 14 321
8 50 1 8 1 1 3 30 1 15 511
Table 2. Customer quintiles and RFM values of customers
RFM provides a simple framework for quantifying customer behavior. For example, it is
possible to infer from Table 2 that customer with id 9, which has RFM score 155, has made a
high number of purchases with high monetary values but not for a long time. Something
might have gone wrong with this customer, for example, he/she has most likely defected to
a competitor's products and services or has found an alternate source and that is why
his/her recency score is low. At this situation, marketers can contact with this customer and
get feedbacks about how to do it better because he/she is one of the valuable customers
according to his frequency and monetary values. Moreover, it is possible to plan a customer
reactivation program and send him/her an extreme promotion in an effort to get his/her

www.intechopen.com
94 Knowledge-Oriented Applications in Data Mining

attention. While customers with score 155 need a reminder, 551's need to be upsold, and
515's need a sticky recurring relationship. For example, if the RFM score of a customer is
identified as 515, marketers can prepare a special customer packet that includes a thank-you
letter, a list of company benefits, and an incentive to make another purchase from the online
store within the next 30 days.
Several studies have discussed the different versions of RFM analysis. For example, in
Weighted RFM (WRFM) version, each R,F,M value is multiplied by a weight value, wR, wF
and wM according to its relative importance to make intuitive judgments about ranking
ordering. Another version, Timely RFM (TRFM) was proposed to deal with the product
periodicity i.e. to analyze different product demands in different times. RFD (Recency,
Frequency, Duration) version was proposed for the web site visitors to consider the duration
i.e. how long someone spends on a website. RML (Recency, Monetary and Loyalty) is an
adaptation of RFM, for annual transaction environments. Loyalty is typically a normalized
form of Frequency in an annual period. RFR (Recency, Frequency, Reach) was proposed for
social graph, i.e. Recency - last post, Frequency - total number of posts, Reach - networks,
friends. FRAT (Frequency, Recency, Amount and Type of goods) is an extended version of
RFM. It induces an improvement of the segmentation by way of taking into account the
categories of bought products, for example, 0 - no buy, 1 - buy a compact car, 2 - buy an
economy car, 3- buy a midsize car, 4 - buy a luxury car, where the order is defined in
increasing order of size.

3. Data mining + RFM

3.1 Clustering using RFM
In recent years, several researchers have considered RFM variables in developing clustering
models. For example, Hosseini et al. (2010) combined weighted RFM model into K-Means
algorithm to improve Customer Relationship Management (CRM) for enterprises. Wu et al.
(2009) applied RFM model and K-Means method in the value analysis of the customer
database of an outfitter in Taiwan to establish strong relationship and eventually
consolidate customer loyalty for high profitable long-term customers. Chuang and Shen
(2008) first assessed the weights of R, F, M in order to know their relative importance by
Analytical Hierarchy Process method, then evaluated Customer Lifetime Values (CLV) by
clustering analysis and finally, sorted customers by self-organizing map method to
recognize high value customer groups.
Differently from the previous Clustering+RFM studies, this chapter proposes using K-
Means++ (Arthur & Vassilvitskii, 2007) algorithm to find customer segments with similar
RFM values. We propose K-Means++ algorithm instead of other clustering algorithms such
as K-Means, self-organizing map because of its advantages in terms of runtime and
clustering quality.
K-Means++ was proposed as a specific way of choosing centers for the K-Means algorithm,
instead of generating randomly. It determines the initial center points by calculating their
squared distance from the closest center already chosen. Through new seeding method, K-
Means++ consistently finds better clusters than K-Means and yields a much faster because
the initialization procedure that ultimately determines the number of iterations to run before
stopping. For example, on a small dataset, K-Means++ terminates almost twice as fast while
achieving potential function values about 20% better, on the larger dataset, it is obtained up
to 70% faster and the potential value is better by factors of 10 to 1000. (Arthur &

www.intechopen.com
Data Mining Using RFM Analysis 95

Vassilvitskii, 2007) For these reasons, we propose K-Means++ algorithm in this chapter,
instead of K-Means or other clustering algorithms.
K-Means++ is a partitioning cluster algorithm by grouping n vectors based on attributes into
k partitions, where k < n, according to some measure. The name comes from the fact that k
clusters are determined and the centre of a cluster is the mean of all vectors within this
cluster. The algorithm starts with determining k appropriate initial centroids, then assigns
vectors to the nearest centroid using Euclidean distance and re-computes the new centroids
as means of the assigned data vectors. This process is repeated over and over again until
vectors no longer changed clusters between iterations.

3.2 Classification using RFM

Recently, integration of classification techniques and RFM was studied by Olson et al. (2009)
to analyze customers response possibilities to a specific product promotion. They compared
three data mining techniques: logistic regression, decision trees and neural networks, and
discussed the relative tradeoffs among these data mining algorithms in the context of
customer segmentation. Cheng and Chen (2009) also combined RFM attributes and rough
set theory (the LEM2 algorithm) to mine classification rules that help enterprises finding out
the characteristics of customers in order to strengthen CRM. Furthermore, in order to
evaluate the accuracy rate of the generated classification rules, they compared their
approach with different three methods: Decision Tree, Artificial Neural Networks and
Naive Bayes. According to the empirical results, their procedure outperforms the other
methods listed in terms of accuracy rate. Ha (2007) used decision tree technique to track
changes in RFM values of customers over time, to discover classification rules related to
transition paths and thus to predict the next customers RFM values from the current
customers RFM values.
Differently from the previous Classification+RFM studies, we apply a classification
algorithm using customer segments discovered by a clustering algorithm and propose the
discovery of classification rules by considering customers demographic variables such as
their ages, genders, occupations, and marital statuses.

3.3 Association rule mining using RFM

In data mining, association rules are descriptive patterns of the form XY, where X is
termed the left-hand-side, and is the conditional part of an association rule; meanwhile, Y is
called the right-hand-side, and is the consequent part. Association rule mining (ARM) is a
task for discovering the hidden, interesting association rules, between items in the database,
having support minsup threshold. The support of an association rule indicates how
frequently that rule occurs in the data. Higher support corresponds to a stronger correlation
between the items in the database.
Several studies applied ARM using RFM variables to analyze customer behaviors. For
example, Chen et al. (2005) recorded all customer behavior patterns (emerging patterns,
added patterns, perished patterns and unexpected patterns) generated by ARM for tracking
changes in customer behaviors at different time snapshots. Liu and Shih (2005) proposed an
approach depending on the idea is that if customers have had similar behavior, then they
are very likely also to have similar RFM values. They firstly applied two hybrid methods
(Weighted RFM-based method and the preference-based Collaborative Filtering method),
and then extracted frequent patterns to represent the common behavior of customers with

www.intechopen.com
96 Knowledge-Oriented Applications in Data Mining

similar purchases. Niyagas et al. (2006) used association rule mining technique and
marketing techniques (RFM analysis) together to analyze historical data of e-banking usages
from a commercial bank in Thailand. They applied Apriori algorithm to detect the
relationships within the features of e-banking services.
Sequential Pattern Mining (SPM) is the extended version of the ARM. While ARM does not
consider the order of transactions, SPM extracts frequent sequences while maintaining their
order. SPM is more complicated than ARM because not only the frequent itemsets but also
the temporal relationships must be found. Recently, SPM and RFM model were studied
together. Chen et al. (2009) developed a novel algorithm for generating all RFM sequential
patterns from customers purchasing data. Liu et al. (2009) proposed a novel hybrid
recommendation method that combines the segmentation-based sequential rule method
with the segmentation-based K-Nearest Neighbors-Collaborative Filtering (KNN-CF)
method. In their proposed method, sequential rules are extracted using customers RFM
values from the purchase sequences in the database.
Differently from the previous ARM+RFM and SPM+RFM studies, this chapter proposes the
application of ARM after clustering and classification tasks to provide better product
recommendations to customers i.e. according to their segments, RFM values and
demographic variables.

4. Integrated approach
This section presents a new three-step approach which uses RFM analysis in data mining
tasks. In our approach, (i) once clustering task is used to find customer segments with similar
RFM values, (ii) then, classification rules are discovered using demographic variables (age,
gender, education level etc.) and RFM values of customer segments to predict future
customer behaviors, (iii) finally; association rule mining is carried out for product
recommendation.
The proposed model can assist managers in developing better marketing strategies that fully
utilize the knowledge resulting from data mining and RFM analysis. It is useful for
predicting customer behaviors according to their demographic variables, because not all
customers have purchased identical amounts, some have ordered more often, and some
have ordered more recently. In addition, it provides better product recommendations than
simple recommendations, by considering several parameters together: customers segment,
the current RFM values of the customer, potential future customer behavior and products
frequently purchased together.
Figure 1 shows the IPO (Input, Process and Output) diagram of the proposed model. The
model consists of five major parts: data preprocessing, RFM analysis, customer
segmentation, prediction, and product recommendation with their evaluation processes.
Each part of the approach is applied one after another. The output of each part becomes the
input of the next part(s). The detail processes of each part are expressed as follows.
Step 1. Data Preprocessing
Data preprocessing step is needed to make knowledge discovery easier and
correctly. Data preparation operations such as reduction in number of attributes,
outlier detection, normalization, discretization, concept hierarch generation
significantly improve the model; in fact a further increasing the prediction accuracy
and saving in elapsed time.
In this step, the following operations should be made:

www.intechopen.com
Data Mining Using RFM Analysis 97

Fig. 1. IPO (Input, Process, Output) diagram of the proposed model

Dimensionality Reduction: Unnecessary attributes should be deleted, such as
attributes that have only a few values (the others are null) or have only single
value.
1.1 Filling: Missing values should be filled in using an appropriate approach.
1.2 Handling: Outliers and inaccurate values should be handled and removed from
the dataset.
1.3 Transformation: Data should be transformed into an appropriate format.
1.4 Discretization: Before association rule mining task, continuous attributes should
be encoded by discretizing the original values into a small number of value
ranges. Because they have nearly a different value for every case; with such a
high cardinality they provide little meaning to the association rule mining
process. One common example of this phenomenon is the attribute that stores
age values. The age attribute can be grouped into four ranges such as child (0-
12), teenager (13-19), adult (20-59) and senior (60+).
1.5 Concept Hierarchy Generation: This method can be used to replace low level
concepts (such as cities Istanbul, Ankara, or Izmir) by higher level concepts
(such as states Marmara, Central Anatolia or Aegean).
Step 2. RFM Analysis
In this step, RFM analysis is applied by defining the scaling of RFM attributes.
This process is divided into four parts introduced in the following:
2.1 Sort the data of three RFM attributes by descending or ascending order.
2.2 Partition the three RFM attributes respectively into 5 equal parts and each
part is equal to 20% of all. The five parts are assigned 5, 4, 3, 2 and 1 score that
refer to the customer contributions. The 5 refers to the most customer
contribution, while 1 refers to the least contribution to revenue.

www.intechopen.com
98 Knowledge-Oriented Applications in Data Mining

2.3 Repeat the previous sub-processes (2.1 and 2.2) for each R-F-M attribute
individually. There are total 125 (5 x 5 x 5) combinations since each attribute in
RFM attributes has 5 scaling (5, 4, 3, 2 and 1).
Step 3. Customer Segmentation
This step divides customers into numerous groups with similar RFM values, and
assigns each customer to an appropriate segment. RFM analysis is used to evaluate
customer loyalty, and thus identify the target customers with high RFM values by
clustering analysis. The main advantage of this process is to be able to adopt
different marketing strategies for different customer segments. Moreover,
clustering customers into different groups improves the quality of
recommendation, helps decision-makers identify market segments more clearly and
therefore develop more effective strategies.
The detail process of this stage is expressed into two sub-steps.
3.1 Clustering: According to RFM attributes for each customer, data is partitioned
into k clusters using the K-Means++ algorithm. (Arthur & Vassilvitskii, 2007)
We propose K-Means++ algorithm instead of other clustering algorithms such
as K-Means, SOM because of its advantages explained in Section 3.1.

A2,...,Ap}, and Ar A, which contains the intervals since last transactions, Af

Let D be a dataset expressed in terms of p attributes from the set A = {Al,

A, which contains the number of transactions within a certain period, and Am

tuple t D has p tuples t = (CustomerID, ri, fi,mi,), where ri Range(Ar) is a

A, which contains the amount of money spent within a certain period. Each

value in the range of the attribute Ar, fi Range(Af) is a value in the range of
the attribute Af , mi Range(Am) is a value in the range of the attribute Am.
Dataset D expressed as D=<(1,r1, f1,m1,), (2,r2, f2,m2,),> is partitioned into
k clusters C = (Cl, C2,..., Ck}.
3.2 Evaluation of Clustering Results: The purpose of this step is to evaluate the quality
of the clusters, to ensure compact clusters with little deviation from the cluster
centroids and while to ensure larger separation between different clusters.
Different methods can be used for evaluating the efficiency of data
segmentation such as Standard Deviation () defined in Eq. 1, Sum of Squared
Error (SSE) defined in Eq.2.

(xi -c)2
N
= N-1
1 (1)
i=1

where xi (i=1,2,..N) is an element in the cluster with N objects and c is the center
of the cluster.

SSE = dist(ci , x )2
K
(2)
i = 1 xci

where k is the number of clusters and ci is the center of ith cluster.

Step 4. Prediction
In this step, classification rules are discovered using demographic variables (age,
gender, education level etc.) and RFM values of customer segments to predict

www.intechopen.com
Data Mining Using RFM Analysis 99

= Aegean then RFM, where the sign denotes that the value is greater than an
future customer behaviors. For example, if age = teenager and gender = male and state

average and sign denotes that the value is smaller than an average.
The rationale of this step is that if customers have similar demographic values, then
they are very likely also to have similar RFM values. In fiercely competitive
environments, discovering classification rules using customer demographic values
is important for helping decision makers to target customer profiles more clearly.
Additionally, the effect of classification rules on recommendations should be
investigated to make more effective marketing strategies.
The detail process of this stage is expressed into two sub-steps.
4.1 Classification: Using customer demographic variables and RFM attributes,
classification rules are discovered by C4.5 Decision Tree (Quinlan, 1993)
algorithm. In data analysis techniques, the capabilities of C4.5 for classifying
large datasets have already been confirmed in many studies.
C4.5 algorithm first grows an initial tree using the divide-and-conquer strategy
and then prunes the tree to avoid overfitting problem. It calculates overall
entropy and information gains of all attributes. The attribute with the highest
information gain is chosen to make the decision. So, at each node of tree, C4.5
chooses one attribute that most effectively splits the training data into subsets
with the best cut point, according to the entropy and information gain.

and k classes from the set C = (Cl, C2,..., Ck}. Thus each sample d D has p+1
Let D be a dataset expressed in terms of p attributes from the set A = {Al, A2,...,Ap},

tuples d = <V1, V2,.., Vp; Cj>, where Vi Range(Ai) is a value in the range of the
attribute Ai A and Cj C. A decision tree is constructed using C4.5 algorithm
that selects an attribute Ai and a subset of its values Vi to branch on.
4.2 Evaluation of Classification Accuracy: Commonly used validation techniques for
classification are simple validation, cross validation, n-fold cross validation, and
bootstrap method. In our model, we propose n-fold cross validation technique
because it matters less how the data gets divided. In this technique, dataset is
divided into n subsets and the method is repeated n times. Each time, one of the n
subsets is used as the test set and the other n-1 subsets are put together to form a
training set. Then the average error across all n trials is computed.
Step 5. Product Recommendation
The core concept of this work is to extract recommendation rules from each
customer group by considering classification rules and using FP-Growth Algorithm
(Han et al., 2000). So, the purpose of this step is to identify the associations between
customer segments, customer profiles and product items purchased together. By
applying such an algorithm, it is possible to recommend products with associated
rankings, which results in better customer satisfaction and cross selling.
The detail process of this stage is expressed into two sub-steps.
5.1 ARM: FP-Growth (Frequent Pattern Growth) is one of the Association Rule
Mining (ARM) algorithms. Among the other ARM algorithms such as Apriori,
Eclat, Mafia, it extracts the rules very fast from data by constructing a prefix
tree and traversing this tree to generate rules. The algorithm scans the database
two times only. Because of these reasons, FP-Growth algorithm is preferred in
this study.

www.intechopen.com
100 Knowledge-Oriented Applications in Data Mining

FP-Growth starts with compressing the database into a frequent-pattern tree

(FP-Tree). During this process, it also constructs a header table which lists all
frequent 1-itemsets to improve the performance of the tree traversal. Each item
in the header table consists of two fields: item name and head of node link,
which points to its first occurrence in the tree. After constructing FP-Tree and
header table, the algorithm starts to mine the FP-tree by considering the items
from the bottom of the header table and by recursively building conditional
FP-Trees.
5.2 Evaluation of Association Rules: ARM algorithms use support and confidence
thresholds and usually produce a large number of association rules which may
not be interesting. An association rule is valid if it satisfies some evaluation
measures. Evaluation process is needed to handle a measure in order to
evaluate its interestingness.
In our approach, we propose to evaluate interestingness of mined rules and to
express the relevance of rules with two descriptive criteria: Lift and Loevinger.
These two criteria are defined on itemsets X, Y and rule R: XY as follows:

P(XY)
Lift(R)= (3)
P(X)P(Y)

P( X )P(-Y )
Loevinger(R)=1 - (4)
P( X - Y )

Lift criterion represents the probability scale coefficient of having Y when X

occurs. Loevinger criterion normalizes the centered confidence of a rule
according to the probability of not satisfying its consequent part Y. In general,
greater Lift and Loevinge values indicate stronger associations.

5. Case study
This section presents a case study which demonstrates how our proposed model was
applied on the real-world data collected by a sports store. All steps of proposed model using
a real world data is expressed in detail.

5.1 Data preprocessing

Dataset used in this case study was provided by a sports store in Turkey and collected
through its e-commerce website within two years period. The complete dataset included
1584 different product demands in 54 sub-groups and 6149 purchase orders of 2666
individual customers. The purchase orders included many columns such as transaction id,
product id, customer id, ordering date, quantity, ordering amount (price), sales type,
discount and whether or not promotion was involved. While customer table included
demographic variables such as age, gender, marital status, education level and geographic
region; product table included attributes such as barcode, brand, color, category, sub-
category, usage type and season.
Data preprocessing step handles outliers, fills missing values and makes dimensionally
reduction, transformation, concept hierarchy generation, normalization and discretization.
From the sport dataset, unnecessary attributes like e-mail addresses, telephone number

www.intechopen.com
Data Mining Using RFM Analysis 101

were obviously inappropriate to be used in data mining and were discarded. Continuous
attributes were encoded by discretizing the original values into a small number of value
ranges. For example, the age attribute was grouped into four ranges: child (0-12), teenager
(13-19), adult (20-59) and senior (60+); the number of children attribute was replaced with
four groups: 0, 1, 2 and 3+. In addition, gender attribute was encoded as m and f instead of
male and female. Furthermore, concept hierarchy generation method was used to replace low
level concepts (city) by higher level concepts (state). Recency attribute was constructed by
calculating time interval between the last transaction date and present for each customer.
Frequency attribute was constructed by finding the number of transactions that each
customer has made within the certain period. Monetary attribute was constructed by
calculating the cumulative total of money spent by each customer. Table 3 shows the partial
data from customers, products and orders tables.
Customers
CID Age Sex State Education Marital S. Child Year
5 Teenager M Aegean Middle NeverM 0 4
8 Adult M Marmara HighSchool Married 0 3
19 Adult F BlackSea HighSchool Married 3+ 4

Products
PID PName Price Brand Group Type Color Sex
100 NK DRI FIT PO 42 Nike TShirt Running NK10 Male
106 PM AIKI JR 81 Puma Sneaker Soccer PM03 Child
110 AD MALV OH 125 Adidas Jersey Soccer AD05 Male

Orders
TID PID CID Date Quantity Discount Total Type
T1 106 19 2008.12.2 1 0 81 SS
T2 100 8 2008.12.2 1 0 42 YS
T3 110 5 2008.12.3 1 0 125 SS

Table 3. An example data from customers, products and orders tables

5.2 RFM model

All customers were ranked by considering their recency, frequency and monetary values
and they were represented by R-F-M codes. Table 4 shows example R-F-M values of some
customers after RFM analysis. For example, it is possible to infer from the first row in Table
4 that customer with id 5 has R-F-M values 4-3-4 respectively. This customer has made a
high number of purchases with high monetary values, not long ago.
Figure 2 shows the distribution of the number of customers with respect to their RFM
values. The distribution of RFM values varies within the limits of 0 - 4.6%. At the most, the
customers have the RFM value 555 (125 customers), followed by RFM value 113 (108
customers), and next, 107 customers have the RFM value 321. Some RFM values such as 121,
125, 231, 311 etc. were not assigned to any customer.

www.intechopen.com
102 Knowledge-Oriented Applications in Data Mining

Recency Frequency Monetary

CID R F M RFM
(Day) (Number) (TL)
5 95 4 237 4 3 4 434
8 269 10 790 5 5 5 555
19 321 1 81 1 1 2 112

Table 4. Example R-F-M values of some customers after RFM analysis

Fig. 2. RFM distribution: 125 possible RFM values and the number of customers

5.3 Customer segmentation

K-Means++ clustering was employed to group customers with similar RFM values.
Customers were segmented into eight target markets in terms of the period since the last
transaction (recency), purchase frequency and total purchase expenditure (monetary). The k

obtained by assigning or , according to the average to R,F,M values of a cluster being less
parameter was set to 8, since eight (2x2x2) possible combinations of inputs (RFM) can be

the overall average R (F, M), then an upward arrow was included, otherwise and
than or greater than the overall average. If the average R (F, M) value of a cluster exceeded

downward arrow was included. For example, RFM represents that the average
recency value of a customer segment is greater than overall average, while frequency and
monetary average values are smaller than overall averages. These eight customer groups
include best customers (most valuable), valuable customers, shoppers, first-time customers,
churn customers, frequent customers, spenders, and uncertain customers (least valuable).
Table 5 presents the result, listing eight clusters, each with the corresponding number of
customers, their average actual and scaled R, F and M values. The last row also shows the
overall average for all customers. The last two columns of Table 5 show the RFM pattern for
each cluster and corresponding customer type. While cluster C5 contains the maximum
number of customers (425 customers, 16%), C6 includes the minimum, only 135
customers (5%).
Customer segment C1 contains the most valuable customers, because it consists of
customers who have recently made regular purchases, and also have higher average

www.intechopen.com
Data Mining Using RFM Analysis 103

purchase frequency and purchase expenditure. It is followed by cluster C2, and next cluster
C3. Cluster C4 (RFM) may include first-time customers, who have recently visited the
company, with higher recency and lower purchase frequency and monetary expenditure.
Customers in C5 have made a high number of purchases with high monetary values but not
for a long time. Something might have gone wrong with these customers, and therefore, it
seems to be an indicator of churn likelihood. It is needed to contact with these customers i.e.
sending an e-mail, and to plan a customer reactivation program i.e. promotion suggestion.
Cluster 8 is concluded to be the least valuable for the business, because customers coded as
111, 112, 121 are generally the least likely to buy again.

Recency Frequency Monetary

RFM Customer
Cluster Size (Avg.) (Avg.) (Avg.)
Pattern Type
Day R # F TL M
C1 309 65.4 4.57 6.28 4.89 485.1 4.79 RFM Best
C2 392 83.5 4.32 1.52 3.44 146.8 3.42 RFM Valuable
C3 415 75.1 4.44 1.18 3.05 70.1 1.49 RFM Shopper
C4 300 202.4 2.86 1.01 2.02 69.5 1.47 RFM FirstTime
C5 425 247.8 2.22 4.27 4.51 387.4 4.67 RFM Churn
C6 135 325.8 1.38 2.26 3.76 137.5 2.94 RFM Frequent
C7 381 290.1 1.86 1.00 1.41 138.1 3.33 RFM Spenders
C8 309 339.1 1.35 1.00 1.00 69.5 1.53 RFM Uncertain
Overall 2666 2.85 3.01 2.95
Table 5. The customer segments generated by K-Means++ clustering based on RFM values
The clusters that have RFM values with at least two upper arrow () can be selected as
target ones, all customers who belong to these clusters become candidates for conducting
suitable marketing strategies, which attract the most attention.
After customer segmentation, standard deviation and SSE metrics were used to evaluate
clustering results. All clusters had a lower standard deviation and SSE values. The result, as
shown in Figure 3, confirmed that these eight clusters were significantly distinguished by
recency, frequency, and monetary. Standard deviation values ranges from 0.67 being the
highest, to 0.33 being the lowest. In the experiments, K-Means++ algorithm was run 10 times
with different initial center values and the clustering result with minimum SSE was selected
as final result.

0.8
Standard Deviation

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
C1 C2 C3 C4 C5 C6
Clusters
Fig. 3. Standard deviations of clusters (customer segments)

www.intechopen.com
104 Knowledge-Oriented Applications in Data Mining

5.4. Customer behavior prediction

A customer segment is not as enough to identify, and then to predict customers behavior.
Many direct marketers believe that the RFM variables of customers are generally associated
with customer profiling. For example, customers with profiles age = teenager and gender =
female and state = Aegean can generally have RFM pattern, while customers with profiles
age = senior and gender = male and state = EasternAnatolia can generally have RFM
pattern. For this reason, in this step, classification rules were discovered using demographic
variables (age, gender, education level etc.) and RFM values of customer segments.
Figure 4 shows a part of classification rules, found in the case study, that identify customer
profiles and the associated RFM values. For example, rule 1 shows that customer profile
with (State=Aegean, EducationLevel=Bachelors, MaritalStatus=Married, Gender=M) is
highly related to RFM pattern. Similarly, classification rule 5 represents that a customer
profile (State=EasternAnatolia, Gender=F) is dominant or most strongly associated with
RFM pattern.

Rule 1: if State=Aegean and EducationLevel=Bachelors and MaritalStatus=Married

and Gender=M then RFM
Rule 2: if State=Aegean and MaritalStatus=Married and Gender=M and Age=Adult
then RFM
Rule 3: if State=Marmara and Membership=3 and EducationLevel=HighSchool and
Children=0 then RFM
Rule 4: if State=CentralAnatolia and Age=Teenager and MaritalStatus=NeverMarried
and Membership=3 then RFM
Rule 5: if State=SouthEasternAnatolia and Children=3+ and Gender=M then RFM
Rule 6: if State=Mediterranean and EducationLevel=Middle then RFM
Rule 7: if State=EasternAnatolia and Gender=F then RFM
Rule 8: if State=BlackSea and Children=3+ and Gender=F then RFM

Fig. 4. A part of classification rules found in the case study

In our experiments, classification accuracy was observed by using 5-fold cross validation
technique. The highest classification accuracy 81% is obtained when different values were
given to parameters (confidence factor, minimum number of objects, number of folds etc.) as
inputs.

5.5 Product recommendation

In the proposed approach, after generating classification rules, association rule mining was
applied to extract recommendation rules, namely, frequent purchase patterns from each
group of customers. The extracted frequent purchase patterns represent the common
purchasing behavior of customers with similar RFM values and with similar demographic
variables. For example, not all women age 45-54 have the same tendency to purchase a
product; so we should also consider their RFM values, customer segments and the other
products frequently purchased together with that product.

www.intechopen.com
Data Mining Using RFM Analysis 105

After customers were classified by demographic variables, the recommendation list was
generated by feature attributes determined using a classification rule inducer. Parameters
were set up to identify association rules that had at least 40% confidence and 2% support
imposed on the FP-Growth association rule algorithm. Figure 5 shows a part of association
rules, found in the case study. For example, if a customer in segment C3 (RFM) buys a
soccer ball, then marketers should recommend backpack and water bottles products.
However, if a customer in segment C4 (RFM) buys a soccer ball, then marketers should
recommend of-kick product. Other rules (Rule 7 and Rule 8) denote that marketers should
recommend two different products (Reebok Sneakers or Converse Shoes) to customers
according to their different RFM values.

Rule 1: {C1, Adidas soccer jersey (man), Adidas soccer jersey (woman)} {Adidas
soccer jersey (child)}
Rule 2: {M>3, Adidas Sneaker (child)} {Adidas Socks, Adidas Equipment Bag}
Rule 3: {C3, Adidas Soccer ball} {Adidas Backpack (unisex), Adidas Water Bottles}
Rule 4: {C4, Adidas Soccer ball} {Nike of-kick}
Rule 5: {C5, Converse Sneaker (woman), Puma Sneaker (man)} {Nike Cap (unisex)}
Rule 6: {C6, Adidas T-Shirt (male)} {Adidas Short (male), Adidas Training Bag}
Rule 7: {R<=3, F<=3, M>3} {Reebok Sneakers}
Rule 8: {R<=3, F<=3, M<=3} {Converse Shoes}

Fig. 5. A part of association rule set on support 2% and confidence 40% for each customer
segment
In the evaluation process, association rules were reduced by more than 50% to the set of
potentially interesting and valuable rules. For example, the number of association rules
related to C4 customer segment was reduced from 67 to 42. These reduction percentages
also give weight to the need of taking into consideration the information brought by the
confirmation property.
In the proposed approach, it is possible to predict the customer segment of a new customer
from classification rules, according to her/his profile, and then a recommendation list can be
generated according to his/her predicted segment.

6. Conclusion
This chapter proposes a novel three-step approach which uses RFM analysis in three data
mining tasks: clustering, classification and association rule mining, applied one after
another. Firstly, customer segments with similar RFM values are identified to be able to
adopt different marketing strategies for different customer segments. Secondly,
classification rules are discovered using demographic variables (age, gender, education level
etc.) and RFM values of customer segments to predict future customer behaviors and to
target customer profiles more clearly. Thirdly, association rules are discovered to identify
the associations between customer segments, customer profiles and product items
purchased, and therefore to recommend products with associated rankings, which results in
better customer satisfaction and cross selling.

www.intechopen.com
106 Knowledge-Oriented Applications in Data Mining

This chapter presents incorporating RFM analysis into data mining techniques to provide
market intelligence. It aims to bring attention of data miners and marketers to the
importance and advantages of using RFM analysis in data mining. In order to evaluate the
proposed model and empirically demonstrate the benefits of using this model in direct
marketing, a case study was carried out using the datasets collected within two years period
by a sports store in Turkey through its e-commerce website. According to experimental
study results, proposed approach provides better product recommendations than simple
recommendations, by considering several parameters together: customers segment, the
current RFM values of the customer, potential future customer behavior and products
frequently purchased together.
Future research can focus in the followings: First, the proposed approach can be tested for
different versions of RFM such as Weighted RFM (WRFM), Timely RFM (TRFM), FRAT
(Frequency, Recency, Amount and Type of goods). As the number of additional variables
increases, the number of cells will geometrically increase. For example, if we add two types
of product parameter, the number of FRAT cells becomes 2 5 5 5 = 500. Thus, it is
unrealistic to estimate RFM model with more than two additional variables. Second, the
effectiveness of the proposed approach can be evaluated for different application domains
such as for the web site visitors (RFD), for annual transaction environments (RML), and for
social graphs (RFR).

7. References
Arthur, D. & Vassilvitskii, S. (2007). K-Means++ The advantages of careful seeding,
Proceedings of ACM-SIAM Symposium on Discrete Algorithms, pp. 1027-1035,
ISBN:978-0898716245, New Orleans, January 2007, Society for Industrial and
Applied Mathematics, USA.
Blattberg, R.C.; Kim, B-D. & Neslin, S.A. (2008). Database Marketing: Analyzing and Managing
Customers, Chapter 12, pp. 323-337, Springer, ISBN: 978-0387725789, New York,
USA.
Bult, J. R. & Wansbeek, T. (1995). Optimal selection for direct mail, Marketing Science, Vol. 14,
No. 4, (Fall 1995) 378-394, ISSN:0732-2399.
Chan, C.C.H. (2008). Intelligent value-based customer segmentation method for campaign
management: A case study of automobile retailer, Expert Systems with Applications,
Vol. 34, No. 4, (May 2008) 2754-2762, ISSN:0957-4174.
Chen, M.; Chiu, A. & Chang, H. (2005). Mining changes in customer behavior in retail
marketing, Expert Systems with Applications, Vol. 28, No. 4, (May 2005) 773-781,
ISSN:0957-4174.
Chen, Y-L.; Kuo, M-H.; Wu, S-Y. & Tang, K. (2009). Discovering recency, frequency, and
monetary (RFM) sequential patterns from customers purchasing data, Electronic
Commerce Research and Applications, Vol. 8, No. 5, (October 2009) 241-251, ISSN:
1567-4223.
Cheng, C-H. & Chen, Y-S. (2009). Classifying the segmentation of customer value via RFM
model and RS theory, Expert Systems with Applications, Vol. 36, No. 3, (April 2009)
4176-4184, ISSN: 0957-4174.

www.intechopen.com
Data Mining Using RFM Analysis 107

Chiu, C-Y.; Kuo, I-T. & Chen, P-C. (2009). A market segmentation system for consumer
electronics industry using particle swarm optimization and honey bee mating
optimization, Global Perspective for Competitive Enterprise, Economy and Ecology,
Springer London, pp. 681- 689.
Chuang, H. & Shen, C. (2008). A study on the applications of data mining techniques to
enhance customer lifetime value based on the department store industry,
Proceedings of the 7th International Conference on Machine Learning and Cybernetics, pp.
168-173, ISBN: 978-1424420964, Kunming, China, July 2008, IEEE.
Ha, S.H. (2007). Applying knowledge engineering techniques to customer analysis in the
service industry, Advanced Engineering Informatics, Vol. 21, No. 3, (July 2007) 293
301, ISSN:1474-0346.
Han, J.; Pei, H.& Yin. Y. (2000). Mining Frequent Patterns without Candidate Generation.
Proceedings of Conference on the Management of Data (SIGMOD00), pp. 1-12,
ISBN:1581132174, Dallas, Texas, United States, May 2000, ACM New York, NY,
USA.
Hosseini, S.M.; Maleki, A. & Gholamian, M.R. (2010). Cluster analysis using data
mining approach to develop CRM methodology to assess the customer loyalty,
Expert Systems with Applications, Vol. 37, No. 7, (July 2010) 52595264, ISSN:0957-
4174.
Kim, H. K.; Im, K. H. & Park, S. C. (2010). DSS for computer security incident response
applying CBR and collaborative response, Expert Systems with Applications, Vol. 37,
No. 1, (January 2010) 852-870, ISSN:0957-4174.
Li, S-T.; Shue, L-Y. & Lee, S-F. (2008). Business intelligence approach to supporting strategy-
making of ISP service management, Expert Systems with Applications ,Vol. 35, No. 3,
(October 2008) 739754, ISSN:0957-4174.
Liu, D-R. & Shih, Y-Y. (2005). Integrating AHP and data mining for product
recommendation based on customer lifetime value, Information & Management, Vol.
42, No. 3, (March 2005) 387-400, ISSN:0378-7206.
Liu, D-R.; Lai, C-H. & Lee, W-J. (2009). A hybrid of sequential rules and collaborative
filtering for product recommendation, Information Sciences, Vol. 179, No. 20,
(September 2009) 3505-3519, ISSN:0020-0255.
McCarty, J. A. & Hastak, M. (2007). Segmentation approaches in data-mining: A comparison
of RFM, CHAID, and logistic regression, Journal of Business Research, Vol. 60, No. 6,
(June 2007) 656-662, ISSN:0148-2963.
Niyagas, W.; Srivihok, A. & Kitisin, S. (2006). Clustering e-banking customer using data
mining and marketing segmentation, ECTI Transaction CIT, Vol. 2, No. 1, (2006) 63-
69.
Olson, D.L.; Cao, Q.; Gu, C. & Lee, D. (2009). Comparison of customer response models,
Service Business, Vol. 3, No. 2, (June 2009) 117-130, ISSN: 1862-8516.
Quinlan, J. R. (1993). C4.5 Programs for Machine Learning, Morgan Kaufmann Publishers. 302
pages.
Swearingen, C. (2009). 101 Powerful Marketing Strategies for Growing Your Business Now!,
SmallBiz Marketing Services, pp. 24-27.

www.intechopen.com
108 Knowledge-Oriented Applications in Data Mining

Wu, H-H.; Chang, E-C. & Lo, C-F. (2009). Applying RFM model and K-Means method in
customer value analysis of an outfitter, Global Perspective for Competitive Enterprise,
Economy and Ecology, ISSN: 1865-5440, Part 12, pp. 665-672, ISBN:978-1848827615,
Springer London.

www.intechopen.com
 Knowledge-Oriented Applications in Data Mining
Edited by Prof. Kimito Funatsu

ISBN 978-953-307-154-1
Hard cover, 442 pages
Publisher InTech
Published online 21, January, 2011
Published in print edition January, 2011

The progress of data mining technology and large public popularity establish a need for a comprehensive text
on the subject. The series of books entitled by 'Data Mining' address the need by presenting in-depth
description of novel mining algorithms and many useful applications. In addition to understanding each section
deeply, the two books present useful hints and strategies to solving problems in the following chapters. The
contributing authors have highlighted many future research directions that will foster multi-disciplinary
collaborations and hence will lead to significant development in the field of data mining.

How to reference
In order to correctly reference this scholarly work, feel free to copy and paste the following:

Derya Birant (2011). Data Mining Using RFM Analysis, Knowledge-Oriented Applications in Data Mining, Prof.
Kimito Funatsu (Ed.), ISBN: 978-953-307-154-1, InTech, Available from:
http://www.intechopen.com/books/knowledge-oriented-applications-in-data-mining/data-mining-using-rfm-
analysis

InTech Europe InTech China

University Campus STeP Ri Unit 405, Office Block, Hotel Equatorial Shanghai
Slavka Krautzeka 83/A No.65, Yan An Road (West), Shanghai, 200040, China
51000 Rijeka, Croatia
Phone: +385 (51) 770 447 Phone: +86-21-62489820
Fax: +385 (51) 686 166 Fax: +86-21-62489821
www.intechopen.com