See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/359864311

Customer segmentation analysis and customer lifetime value prediction using

Pareto/NBD and Gamma-Gamma model

Conference Paper · November 2021

CITATIONS READS

0 150

3 authors, including:

Van-Ho Nguyen Thanh Ho

University of Economics and Law (VNU-HCM) Ho Chi Minh City University of Economics and Law
13 PUBLICATIONS 31 CITATIONS 31 PUBLICATIONS 75 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Thanh Ho on 15 June 2022.

The user has requested enhancement of the downloaded file.

 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Customer segmentation analysis and customer lifetime value

prediction using Pareto/NBD and Gamma-Gamma model
Kim-Giao Tran1,2, Van-Ho Nguyen1,2, Thanh Ho1,2,*
1
University of Economics and Law, Ho Chi Minh City, Vietnam
2
Vietnam National University, Ho Chi Minh City, Vietnam
*Corresponding author: thanhht@uel.edu.vn

ABSTRACT
Customer segmentation helps the organization to manage customer relationships expertly and
gain a deep understanding of customers. With the proliferation of machine learning methods,
this study performs data science algorithms into traditional marketing such as Recency,
Frequency and Monetary (RFM) model, or RFM model with K-Means clustering for customer
segmentation, and Pareto/Negative binomial distribution (NBD), Gamma-Gamma model for
predicting customer lifetime value (CLV). This study experiments on a dataset of 121,317
historical transactions of the bicycle retail industry, including individual customers and
retailers. By the retention analysis, it shows that customers mostly prefer coming back after at
least three months as bicycles are long-termed usable. The combination of segmentations from
RFM models and the CLV results of each customer can show that whether the customers, who
are in well-evaluated segments, have the proportionally high value to the business or not. This
can also be seen as a basis for cross-checking the completeness of these techniques. The
evaluation metrics with high accuracy in training and evaluating the Pareto/NBD and Gamma-
gamma models show that the CLV are well-trained before being compared to the RFM
segmentation. This study confirms the connection between the traditional RFM, the RFM model
with K-Means, and CLV techniques as the results show that customers in good segments, have
high predicted CLV as well. Based on the empirical results, the proposed research models can
be applied in other businesses that will help them get the effective business strategies for each
customer group depending on their financial and human potential.
Keywords: RFM model, customer segmentation, clustering, CLV, Pareto/NBD, customer retention

1. Introduction
In the intense competition and complexity of the business environment, customer
segmentation helps the marketing departments easily define the pivotal solution to attract each
group of customers. Based on the data segmentation, customers are classified into different
groups according to distinguishing similarities such as gender, age, income, products of interest,
and purchasing behaviors (Anitha & Patil, 2019). These characteristics are analyzed and
categorized based on the historical purchasing data of the business. Recency, Frequency, and
Monetary (RFM) has been very famous in marketing as a tool to identify a company’s best
customers by calculating and analyzing their spending habits. RFM analysis weights customers'
importance by scoring them in three measurements such as how recently they have made a
purchase (Recency), how often they have bought (Frequency), and how much they have spent
(Monetary) (Thanh & Son, 2021).
Besides using RFM for customer segmentation, customer lifetime value (CLV),
retention rate and churn rate are a combination of robust metrics to measure customer
satisfaction. While CLV is the discounted value of future profits that the customer spends on
the company (Glady et al., 2009), the retention rate shows the ability of a company to keep its

296 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

existing customers (Ismail et al., 2015). In contrast, the churn rate is the percentage of customers
moving out of a cohort over a particular period.
As Kotler and Keller described customers’ churn as a phenomenon that results in a waste
of money and efforts (Kotler & Keller, 2006), choosing to focus on retaining old customers and
turn them from potential customers into loyal customers will help businesses reduce more costs
than building advertising campaigns to attract new customers. However, the problem is that
when a business has a lot of customers and all of them have made many transactions with the
business, it is tough to know if they are still attached to the business or not. Besides, businesses
cannot calculate precisely when a customer will leave but can only predict based on
probabilities, so this is even more difficult.
While the studies above mainly focused on the RFM model or the Pareto/NBD and
Gamma-Gamma models only. This study identifies the goal of combining the techniques to find
the relationship between them, and desires to provide the managers a multi-dimensional view
of their customers. For more particular, this study wants to detect the bonds between the
traditional RFM segmentation and the K-Means clustering with RFM. Then applying the CLV
results from Pareto/NBD and Gamma-Gamma models to have a more precise insight of the
customers’ value of each group. The relationship between RFM and CLV is also included. That
makes it easier for managers to decide whether to implement appropriate marketing strategies
for each customer group as well as to assess whether existing customer care policies are still
appropriate for retaining customers or not. Besides, this study also first briefly uses the retention
analysis to find out the customers’ purchasing behaviors and patterns before going straight to
the further analysis.
The following content of the article is Section 2, including the theoretical basis and
related studies, to identify models and algorithms suitable for the set goals. Section 3 is the
methodology that describes relevant issues and experimental processes. After the experimental
process, the results and discussion of the identified customer segments are mentioned in Section
4. The last Section is the conclusion and implications of the study.
2. Theoretical background and related work
This section provides the literature overall and some related researches based on the
purpose of this study.
The RFM model is usually used to classify customers and define their behaviors. RFM
records the customers’ transactions under three factors:
(1) Recency is the distance between the last purchasing date of that customer and the
date of implementing the model;
(2) Frequency is the total transactions of that customer;
(3) Monetary is the actual money that the customer had spent on businesses’ products
or services.
The most well-known clustering methods by RFM are customer quintiles (Miglautsch,
2000) and clustering by K-Means.
Clustering using the K-Means algorithm is a method of unsupervised learning used for
data analysis. It generates k points as initial centroids randomly, with k is chosen by users. Each
point is assigned to the cluster with the closest centroid. Then the centroids are updated by
taking the mean of the points of each cluster (Anitha & Patil, 2019; Ismail & Dauda, 2013;
Madhu et al., 2010). The data points may move to different clusters after each iterative
approach. The chosen centroids are defined when there are no point changes clusters or the
centroids remain. The algorithm uses mainly Euclidean distance to measure the distance

297 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

= ( 1, 2, 3, … , m) and =
between data points and centroids (Dwivedi et al., 2014). The formula to calculate Euclidean

( 1, 2, 3, … , m) is described as equation (1):

distance between two multi-dimensional data points

( , )= ( 1 − 1 )2 +( 2 − 2)2 + ⋯+ ( m − m )2 (1)
Although K-Means is the most common algorithm to classify clusters, it still has some
drawbacks. Because the centroids are first chosen randomly, the results can turn out different
for different runs. Besides, defining the right number of clusters is also a tremendous problem
to deal with. Thanh and Son (2021) used the Elbow method to find the optimal number of
clusters then using the Silhouette method to re-evaluate the results above, while Anitha and
Patil (2019) only used the Silhouette score to find the optimal k. These studies pointed out the
efficiency of the clustering method in Data Science and also performed the clustering results in
RFM analysis and provided customers’ different behaviors in specific clusters.
The Elbow method is used to determine the number of clusters of a dataset by using the
visual technique. The graphic obtained the results from the Sum Squared Error (SSE)
calculation, which measures the difference between points in clusters. The more the number of
clusters k, the smaller the SSE value will be. If the value of the former cluster and the value of
the later cluster draw an angle between them, the cluster at the elbow flexion point will be the
chosen cluster or the cluster with the biggest reducing value compared with its former will be
chosen (Thanh & Son, 2021; Humaira & Rasyidah, 2020; Nainggolanet et al., 2015). The
formula of SSE calculation is described as equation (2):
! j

= ( ij, i) 2
(2)

"

Where m is the centroid of the data point x and k is the number of clusters. The graphic
which obtained the values of the SSE calculation for the different number of clusters will
perform the visual looks as an elbow arm. The Elbow method is easy to implement and
adequately fitted with perplexing, huge data, but its weakness is the user must choose the
number of clusters based on experience (Humaira & Rasyidah, 2020).
Along with the Elbow method, the Silhouette score is also an effective way to see how
well each cluster is separated from the others. In the two studies (Anitha & Patil, 2019; Humaira
& Rasyidah, 2020), the authors give two different theories about the range values of the

range #−1, +1$, if it is scored near +1, the clustering quality performed well, if it is valued at
Silhouette score. After researching more deeply, the Silhouette score is informed to be in the

0, we can say there is no distinction between the clusters, and if it is near -1, the clusters were
not distributed well (Ogbuabor & Ugwoke, 2018). The formula to calculate Silhouette score is

0i − 1i
%&ℎ()*++* -.(/*i =
written as equation (3):

max (1i, 0i) (3)

With a is the average intra-cluster distance (the mean distance between i and the data
points in the same cluster), and b is the average inter-cluster distance (the mean distance
between the data point i to all the data points outside its cluster).
Pareto/negative binomial distribution (NBD) model is one of the most classic used RFM
models to calculate CLV. The model mostly used the recency, frequency, and length of the
customer’s observation period to predict the customer’s future purchases (Hellerslia & Talal,
2020). The Pareto/NBD model is developed by Fader and Hardie. They also described the
model that is based on five assumptions (Fader et al., 2004):

298 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

(1) The transactions made by a customer in a period of length + follow a Poisson distribution
with transaction rate 4. It means that they can purchase randomly whenever they want
in their active period, but the rate (in a unit time) is constant.

shape parameter / and scale parameter 5 .

(2) Heterogeneity in transaction rates across customers follows a gamma distribution with

(3) Each customer has an unobserved lifetime + . In other words, the point at which the

6.
customer becomes inactive or churned is distributed exponentially with the dropout rate

shape parameter - and scale parameter 7 .

(4) Heterogeneity in dropout rates across customers follows a gamma distribution with

(5) Each customer has a varied transaction rate 4 and the dropout rate 6 .
The Gamma-Gamma model is the extension of the Pareto/NBD model. While the
Pareto/NBD model only focuses on the recency and frequency factors, the Gamma-Gamma
model uses the monetary component to predict the average future purchasing value (Avinash et
al., 2019; Hellerslia & Talal, 2020). The Pareto/NBD and Gamma-Gamma models are a
powerful combination to calculate CLV. While Pareto/NBD predicts future purchases, the
Gamma-Gamma model allows us to assign a monetary value to each of those future purchases.
To ensure to have the best estimated CLV, these models can be evaluated in the holdout period
before making forecasts.
3. Methodology and proposed research model
Figure 5 describes methodology and proposed research model with three main stages:
(1) Stage 1 is customer segmentation analysis according to RFM. From the input data is a
dataset extracted from the sales department of Microsoft's Adventure Works sample
data, perform data preprocessing, and calculate recency, frequency, and monetary
values for use in the RFM model. After preprocessing the data and realizing the
difference between the data points, the study will implement the standardization for
input data, then using some methods related to K-Means to find out the optimal number
of clusters for segmentation;

Figure 5: Overview of the proposed research model

(2) Stage 2 is to use Pareto/NBD and Gamma-Gamma model to predict the number of
purchases and revenue that customers yield in the future, the root of these two models
is to be exploited and developed from the RFM model. Build the above two models on
299 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

the training set and re-evaluate the predictions on the test set to see the accuracy of the
model. Repeat this loop by changing the indexes in the model until the model gives the
most optimal results;
(3) Stage 3 is from the two optimal models above, performing customer lifetime value
(CLV) prediction.
4. Experimental result and discussion
4.1. Customer segmentation using RFM
The first stage of the experimental process including preprocessing data standardization,
RFM data construction, and K-Means customer segmentation (Figure 5).
4.1.1. Dataset and data preprocessing
The study uses a dataset of customer transactions extracted from the dataset of the
company Adventure Works Cycles. This is a multinational company that manufactures and
sells bicycles to the North American, European, and Asian markets. The extracted dataset
records 121,317 transactions of the company from 06/2011 to 07/2014. This includes both
individual customers and retailers. To analyze the optimal customer segment for each different
market, the study filtered out the transactions made in the US (United States) market for use in
further analysis.
4.1.2. Customer retention analysis
Before jumping into clustering customers by the RFM model, the study will briefly
analyze the company's customer retention situation to find out the insight of its business status.
Adventure Works is a business that manufactures and sells bicycles for both individuals and
resellers, but bicycles are non-essential and can be used in the long term, the number of
customers who have one transaction only over 3 years is very high, at 74.31% (Figure 6).
Meanwhile, the number of customers who used to repeat transactions with the company only
accounted for 25.69% but brought even higher revenue than the others over time. In particular,
there would be a sudden increase in the revenue that this group of customers brought to the
business every 1 month.

Figure 6: Sales by customers over time

It can be seen in Figure 7 that, in the period from 05/2011 to 06/2013, approximately
two years, the number of regular customers was higher, almost all customers returned to make
transactions again with the company. However, starting from 07/2013, when the business had
a sudden growth in attracting more customers, the number of customers leaving when they only
transacted once with the business was very high.

300 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Figure 7: Number of churn and repeated customers over time

Grouping customers according to cohort, also known as grouping customers according
to the timeline from the customer's first transaction (Croll & Yoskovitz, 2013). The formula to

) 8) 0*/ ( .)-+( */- 1&%9* %8 *1.ℎ (8+ℎ

calculate the retention rate is described as equation (4):

*+*8+%(8 /1+* =
) 8) 0*/ ( %8%+%1& .)-+( */-
(4)

The retention rate of each cohort is shown on the horizontal axis of Figure 8. With the
analysis of customer retention rate using the heat chart, it can be seen:

Figure 8: Retention rates in cohort analysis

Customers of the business did not transact regularly once a month, but on average,
customers came back every 2-3 months. With the group of customers having transactions from
the beginning of the observed period, from June 2011, only 4% of customers returned to transact
in the next month. However, with a cycle every 2-3 months, the customer retention rate of the

301 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

business at this time was very high, in the 34th month, it still maintained 67% of the total number
of original customers.
In contrast, the retention rate for new customer groups decreased considerably.
Generally, the company's customer retention policy was appropriate for the period before 2012
and was able to retain this group of loyal customers until the end of the period. However, it
seemed to be no longer suitable for new customer groups, especially when the business in later
period promoted marketing and attracted more customers but cannot keep them. Businesses
should focus more on customer care policies as well as targeted marketing campaigns to attract
returning customers.
4.1.3. Customer segmentation based on RFM scores
This is the traditional and the simplest way that can explain how the RFM model works.
The RFM model is famous for transforming transactional data, which basically includes
CustomerID – unique customer code, SalesOrderID – unique invoice code, ProductID – unique
product code, InvoiceDate – date of the transaction, Quantity – quantity of purchased items,
Unit Price – the price of one item, Country – country of the transaction, into profitability scores
(Zaki et al., 2016). After calculating recency, frequency, and monetary for RFM analysis, the
characteristics of the statistical distribution of these factors such as average, minimum value,
maximum value, as well as quartiles are described in Table 8. The average last purchased date
is 206 days ago with nearly 1.5 purchases and 1473.8 revenue in total.
Table 8
Quartiles description in RFM table
Recency Frequency Monetary
Mean 206.377101 1.466626 1473.809070
Min 0.000000 1.000000 1.374000
Max 1122.000000 12.000000 58662.190608
1st quartile 91.000000 1.000000 21.490000
nd
2 quartile 177.000000 1.000000 69.990000
rd
3 quartile 277.000000 2.000000 2294.990000
While the authors in (Zaki et al., 2016) ranked customers in quintiles. This study chose
to rank them based on the quartiles. Following the related works, customers with the most
recency value will have a 1 R score. In contrast, the ones with the lowest recency received a 4
R score because the customers with more recent transactions are considered more valuable to
the business. This step was implemented repeatedly for the frequency and monetary but in a
reverse way, which means the highest frequency and monetary received 4 scores and the ones
with the lowest had 1 score. Noted that the customers’ value is proportional to RFM scores. By
mapping the RFM scores, we had the worst valuable customers who had an overall RFM score
of 111. On the contrary, the ones with an RFM score of 444 were considered as top customers
to the company.
This study divided customers into segments based on the exemplary segmentation of
the RFM scores, which Jasmin described as a graphic in her blog (Jasmin, 2020). This study
used the exemplary in the reverse RFM scores.

302 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Figure 9: Customer segmentation distribution

Figure 9 shows the RFM segmentation labeling results in a treemap. Based on different
segments, the company needs a specific strategy to develop its business status. The company
had a huge number of new customers as Unsteady Customers (36.53%) with high monetary
value. It is advised to build a long-term relationship with these customers by cross-selling
strategy or specific promotions. Besides, the Customers At Risk segment, which accounted for
16.95%, is also a potential group from which the business can exploit. With a very high
monetary value but having stopped trading for a long time, finding a way to contact and pull
these customers back will bring a great benefit to the business. Top Customers and Active
Customers accounted for a small percentage but the profits that they brought were considerable,
the company cannot lose them. Finally, the company had quite many Inactive and Lost
customers. The study mentioned that the company with main products such as long-term usable
bicycles could have many churned customers, but the managers could research more insight
into these customer groups to find the exact churned reason to re-engage these customers as
much as possible.
4.1.4. Data standardization
After preprocessing the data and preparing the input data for the RFM model with the
corresponding recency, frequency, and monetary values, it was found that there is a huge
difference between these three values, which can affect the model run time and the accuracy of
the algorithms. The study conducted data standardization according to the standard distribution
method (Standard Score), also known as Z-Score (Ismail & Dauda, 2013) to bring the data to a
distribution range that where the mean value of the observations is 0 and the standard deviation

−6
is 1. The formula for standardizing the data is described as equation (5):
:
=
;
(5)

With x is the initial value before standardization, μ is the mean value of the observations,
and σ is the standard deviation of the observations. After standardizing the data, it can be
concluded that the three current recency, frequency, and monetary values weight equally when
included in the analysis in the K-Means clustering model.
303 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

4.1.5. The optimal number of clusters for K-Means

As having described these models’ literature in the theoretical basis part, this study uses
the Elbow method and Silhouette score to find the most optimal number of clusters of the
dataset. The result for the number of clusters from 2 to 9 is described in Figure 10. The graphic

< = 3 or < = 4. Silhouette score will be implemented to re-evaluate the quality of finding
has the visualization as an elbow and the SSE line shows that the elbow flexion point is around

optimal k in the Elbow method to get the final result.

Figure 10. Elbow method result Figure 11. Silhouette score result

< = 3 with 0.74 score is the highest among other clusters. The indicator means that with < =
Figure 11 illustrates the results that Silhouette scored each cluster. It can be seen that

3, the distance from data points to their centroid in each cluster is optimized and the cluster
eccentricity barely occurs. Therefore, the study will use the number of < = 3 to cluster
customers into different levels based on three factors of the RFM model (Recency, Frequency,
and Monetary).
The number of customers and the average value of recency, frequency, and monetary of
each cluster after being divided into 3 clusters by K-Means are all described in Table 9. It can
be seen that the Gold cluster includes the least number of customers who have the most
transactions, relatively recent purchases and bring the highest revenue for the business. In the
other two groups, they have quite similar frequency and monetary mean value, but the average
recency value of one group is nearly twice more than the other one. The group with the most
recency value is labeled as the Bronze group because of without recent transactions with the
business.
Table 9.
Each cluster description
Cluster Customers Mean recency Mean frequency Mean monetary
Gold 216 149.527778 8.606481 11299.873848
Silver 4665 148.998928 1.231726 1162.314786
Bronze 3329 290.471012 1.332532 1272.754954
Figure 12 describes the clustering result with < = 3 in three-dimensional space. The
Silver level is the group with the highest convergence, but there is still confusion between the
data points in the Silver and Bronze levels. The Silver group contains customers who have more
stable recency and frequency indexes while the ones in the Bronze level have even higher
monetary value but stopped trading for a long time. Besides the Gold group with its distinction
from the others, Silver and Bronze were labeled mainly based on the average recency value.

304 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Figure 12: Clustering result by RFM level

4.2. Predicting CLV using Pareto/NBD and Gamma-Gamma model
The second stage of the experimental process including Pareto/NBD and Gamma-
Gamma model data construction (Figure 5), calibration and holdout dataset divided for
evaluation, then using the models to predict CLV.
4.2.1. Constructing input data for Pareto/NBD and Gamma-Gamma model
Because the Pareto/NBD and Gamma-Gamma models use the RFM basis, the data used
for these models are quite the same as the previous data construction. The difference is

with only one transaction have /*A)*8. = 0 and (8*+1/ = 0 in this model. The
Pareto/NBD only considers customers with repeated transactions, which means the customers

Pareto/NBD model also uses another factor, which is the customer lifetime (T) calculated as
the distance from the customer’s first purchase date to the model implementation date.
The data for the Gamma-Gamma model is the same as the data of the Pareto/NBD model
but only the rows have frequency and monetary bigger than 0.
4.2.2. Calibration and holdout dataset
The calibration dataset starts at the beginning of the observed period from 07/06/2011
to 07/07/2013, and the holdout period spans from 08/07/2013 to 07/07/2014, exactly 365 days.
The percentage is approximately 70% in calibration and 30% in the holdout dataset.
4.2.3. Predicting future purchases using Pareto/NBD model
Figure 13 illustrates the number of purchases in the calibration dataset on the x-axis and
the corresponding average number of purchases in the holdout dataset on the y-axis. As can be
seen, the model predicted that the customers with the higher purchases in the calibration would
also have higher purchases in the holdout, except for a slight reduction in customers with 5
purchases in the calibration. In contrast, in actual data, the holdout set shows more
unpredictable volatility.

305 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Figure 13: Actual and predicted purchases of Pareto/NBD model in holdout dataset
Evaluating models can be the most important step in Data Science. This study used
some indicators to evaluate the quality of the prediction model. The formulas of these indicators
are described as equations (6), (7), and (8) (Chicco et al., 2021):
E
1
BC = | − i|
(6)
i
"
E
1
B = ( i − i )2
"
(7)

E
1
B =F ( i − i)2 (8)
"

With X is the actual values and Y is the predicted values. MAE is a measurement of
errors between actual and predicted observations. MSE measures the average of squares of the
errors or can be understood as the average squared difference between the actual and estimated
values. RMSE is also a measurement to evaluate the differences between two observations. The
more these indicators move near to 0, the fewer errors the predictions have. Table 10 shows
that the Pareto/NBD predicted the future purchases fairly well while the evaluation values are
all small, nearly 0.
Table 10
Pareto/NBD purchases prediction evaluation
Types Results
Mean Absolute Error (MAE) 0.7904071127295603
Mean Squared Error (MSE) 0.8850164704192321
Root Mean Squared Error (RMSE) 0.9407531399996665
4.2.4. Predicting the future average order value using the Gamma-Gamma model
The estimated results of the Gamma-Gamma model are shown in Figure 14. The
histogram plots the monetary value distribution of the actual and estimated observations. It

306 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

shows that predicted results tend to be smaller than reality and both predicted and actual
monetary values are concentrated near zero.

Figure 14: Actual and predicted of the Gamma-Gamma model in the holdout set
Because the difference in monetary values is much larger than the factor used in the
previous model. Instead of using metrics such as MAE, MSE, and RMSE, which are often used
for normalized, standardized datasets or have values close to zero, here the chosen option is
dividing monetary values into 5 bins according to ordinal variable and K-Means. Then use the
confusion matrix and the F1-score to evaluate the accuracy of the model.
The confusion matrix Figure 15 shows that the Gamma-Gamma model worked well in
the holdout set while the predictions were mostly divided into the right bins. The F1 score is
0.9 which means the estimation had high accuracy.

Figure 15: Confusion matrix of actual and predicted monetary value

After training and evaluating the models to check their quality, the Gamma-Gamma
model was implemented again in the initial dataset to check if it had the overfitting problem or
not. Fortunately, the model also performed well in the original data set, proved when Figure 16
shows that the actual and predicted monetary values have a linear correlation, and the histogram
gives the results of the true and predicted values almost overlapping (Figure 17). After training
and fixing the Pareto/NBD and Gamma-Gamma model, finding that the two models worked

307 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

quite well and the evaluation was very high, the study would apply these two models to predict
the CLV for the company's customers.

Figure 16: Scatter plot of actual and Figure 17: Histogram of actual and predicted
predicted monetary value in initial dataset monetary value in initial dataset
4.2.5. Predicting CLV values

Figure 18: Average predicted CLV by customer segmentation

Among the 5 groups containing repeated customers of the business in Figure 14, the
model predicted that the Top Customers group has the largest customer lifetime value. The
Active Customers and Unsteady groups have a fairly low value because they did not make many
transactions with the business during the observed period. Although the Customers At Risk
group had not traded with the business for a long time, the number of orders and the amount of
revenue that this customer group could bring is very large for the business, its estimated CLV
is fairly high.
4.3. Discussion
The study had found out the relationship between the original RFM customer
segmentation and the RFM customer clustering by K-Means. Figure 19 describes the total
number of customers in 8 segments divided by RFM score and 3 clusters classified by K-Means.
As can be seen that the customers in Gold level were mostly divided into the Top Customers,
Emerging Customers, and Customers At Risk, these were also the three which had the most
predicted CLV in the previous analysis. This shows that the models used in the study are closely
related to each other.

308 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Besides, as mentioned above, the Bronze and Silver groups had almost the same
Frequency and Monetary indexes, only a big difference in Recency. Therefore, the K-Means
model for clustering has not been fully effective. Since each run of K-Means gave different
clustering results, and the user had to base on those results and label the clusters for each
customer group, this requires a lot of expertise in the field to be able to effectively cluster and
label the group thoroughly. However, the K-Means clustering and the predictions of the
Pareto/NBD and Gamma-Gamma model were matching when the 3 top customer groups in
CLV including customers in Gold level, and the two next most CLV groups contained mostly
customers from the Silver group.

Figure 19: The number of customers by segmentation and RFM level

5. Conclusions and implications
By combining marketing and business knowledge with information technology, a
clearer view of the Adventure Works company was realized. RFM is easy to apply and flexible
method for implementing customer segmentation. As Mark Patron commented that RFM did
not provide the company the profitability and the potential of a customer (Patron, 2004), using
the combination of RFM and CLV results to find hidden potential customers in the business
will be very profitable. Managers can base on that to implement customer care policies such as
discounts and customer gratitude programs for Gold and Silver customers or use cross-selling
strategy to maximize profits from existing customers as well as attract new customers. The
model in this study was defined to use most effectively for this dataset. They can be developed
to use particularly based on the company’s needs.
References
Anitha, P. & Patil, M. M. (2019). RFM model for customer purchase behavior using K-Means
algorithm. Journal of King Saud University – Computer and Information Sciences, 1-8.
doi:10.1016/j.jksuci.2019.12.011
Avinash, A., Sahu, P. & Pahari, A. (2019). Big Data Analytics for Customer Lifetime Value
Prediction. Telecom Business Review, 12(1), 46-49. Retrieved from
http://publishingindia.com/tbr/

309 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Chicco, D., Warrens, M. J. & Jurman, G. (2021). The coefficient of determination R-squared is
more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis
evaluation. PeerJ Computer Science, 7(3), e623. doi:10.7717/peerj-cs.623
Croll, A. & Yoskovitz, B. (2013). Use the Data to Build a Better Startup Faster (1 ed.).
Cambridge: O'Reilly Media.
Dwivedi, S., Pandey, P., Tiwari, M. S. & Kalam, A. (2014). Comparative Study of Clustering
Algorithms Used in Counter Terrorism. IOSR Journal of Computer Engineering (IOSR-
JCE), 16(6), 13-17. doi:10.7763/IJKE.2015.V1.18
Fader, P. S., Hardie, B. G. S & Lee, K. K. (2004). “Counting Your Customers” the Easy Way:
An Alternative to the Pareto/NBD Model. Marketing Science, 24(2), 275-284.
doi:10.1287/mksc.1040.0098
Glady, N., Baesens, B. & Croux, C. (2009). A modified Pareto/NBD approach for predicting
customer lifetime value. An International Journal, 36(2), 2062-2071.
doi:10.1016/j.eswa.2007.12.049
Hellerslia, G & Talal, D. (2020). Comparison of classical RFM models and Machine learning
models in CLV prediction. Norway: Master of Science.
Humaira, H. & Rasyidah, R. (2020). Determining The Appropiate Cluster Number Using
Elbow Method for K-Means Algorithm. Proceedings of the 2nd Workshop on
Multidisciplinary and Applications (WMA) 2018 (pp. 24-25). Padang: EAI.
doi:10.4108/eai.24-1-2018.2292388
Ismail, M. & Dauda, U. (2013). Standardization and Its Effects on K-Means Clustering
Algorithm. Research Journal of Applied Sciences, Engineering and Technology, 6(17),
3299-3303. doi:10.19026/rjaset.6.3638
Ismail, M., Bawa, M. & Safrana, M. J. (2015). Impact Of Marketing Strategy On Customer
Retention In Handloom Industry. 5th International Conference, SEUSL, (pp. 16-25). Sri
Lanka.
Jasmin. (2020, November 12). Machine Learning In Customer Segmentation With RFM-
Analysis. Retrieved from Nextlytics: https://www.nextlytics.com/blog/machine-
learning-in-customer-segmentation-with-rfm-analysis
Kotler, P. & Keller, K. L. (2006). Marketing Management (12th ed.). New Jersey: Pearson
Prentice Hall.
Madhu, Y., Pathakota, S. R. & Srinivasa, T. M. (2010). Enhancing K-means Clustering
Algorithm with Improved Initial Center. International Journal of Computer Science and
Information Technologies, 1(2), 121-125.
Miglautsch, J. R. (2000). Thoughts on RFM scoring. Journal of Database Marketing &
Customer Strategy Management, 8(1), 67-72. doi:10.1057/palgrave.jdm.3240019
Nainggolan, R., Perangin-angin, R., Simarmata, E. & Tarigan, F. A. (2015). Improved the
Performance of the K-Means Cluster Using the Sum of Squared Error (SSE) optimized
by using the Elbow Method. Journal of Physics: Conference Series, 1-6.
doi:10.1088/1742-6596/1361/1/012015
Ogbuabor, G. & Ugwoke, F. N. (2018). Clustering Algorithm For A Healthcare Dataset Using
Silhouette Score Value. International Journal of Computer Science & Information
Technology (IJCSIT), 10(2), 27-37. doi:10.5121/ijcsit.2018.10203

310 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

Patron, M. (2004). Applying RFM segmentation to the SilverMinds catalogue. Journal of

Direct Data and Digital Marketing Practice, 5(3), 269-275.
doi:10.1057/palgrave.im.4340243
Thanh, H. T., Son, N. T. (2021). An interdisciplinary research between analyzing customer
segmentation in marketing and machine learning method. Sci. Tech. Dev. J. - Eco. Law
Manag, 6(1), 2005-2015. doi:10.32508/stdjelm.v6i1.850
Zaki, M., Kandeil, D., Neely, A. & McColl-Kennedy, J. R. (2016). The Fallacy of the Net
Promoter Score: Customer Loyalty Predictive Model. UK: Cambridge Service Alliance.

311 | I C B 2 0 2 1
 The 4th International Conference on Business
Ho Chi Minh City Open University, HCMC, Vietnam. 04 - 05 November 2021

From service recovery to post-recovery customer satisfaction: A

review the role of customer control and transparency
Luu Quang Minh1, Nguyen Ngoc Duy Phuong2,*
1,2,*
International University, Vietnam National University Ho Chi Minh City
*
Corresponding author: nndphuong@hcmiu.edu.vn

ABSTRACT
Service business has trouble retaining their customers because of poor complaints handling and
underestimation of service recovery process. This research paper aims to describe the
relationship between service recovery and customer satisfaction and suggest potential service
recovery implications to increase customer satisfaction. This research is conducted by
comparing different service recovery definition from different researchers’ point of view and
identify common trends and patterns on customers perception of justice, hence, giving support
ideas that can help increase customer satisfaction. The outcome of this study will help service
managers to understand service recovery and service recovery procedure to recover satisfaction
and retain customers. This research paper lacks analyzing the difference in post-recovery
satisfaction difference in age groups, genders, and ethnicities, as well as the difference between
group service failure and individual service failure.
Keywords: Customer satisfaction, service recovery, service failure.

1. Introduction
Service recovery is a term that appears following service failure, as service failure
commonly exists when delivering service to customers. Service failure appears suddenly in the
service delivery process because the process contains multiple steps that involve not only the
human resource of the firm, but also with the help of their own customers. Therefore, service
failure occurs when firms are not able to maintain a consistent quality level, business does not
meet customers’ expectations, and is unavoidable (Gustafsson, 2009; Johnson et al., 2002;
Krishna et al., 2011; Pranić & Roehl, 2013). The incident raises a concern about customers’
dissatisfaction, negative words-of-mouth (WOM), and customers preservation for service
firms. Service firms usually does not notice the long-term effect of losing their customers as
they think that they can fill in the gap of previous customers with new ones. A study in 2006
by Ang and Buttle shows that keeping loyal customers provide a significant financial benefit
such as reducing cost of acquiring new customers, rather than trying to acquire new customers.
However, service failure will not cause customers to go away to their competitors, but
dissatisfaction will (Cheng et al., 2019). Service recovery is a tool to communicate and build
stronger relation with customers in the form of interaction, empathy, and action (Krishna,
Dangayach, & Sharma, 2014). In the same research, service recovery provides customers with
confidence and confirmation, aim at achieving an immediate and long-term satisfaction and
relationships. Other researchers (Baron, Harris, Elliott, Reynolds, & Harris, 2005; Mansori et
al., 2014, as cited in Cheng et al., 2017) emphasizes the importance of service recovery because
of its usefulness in increasing customer satisfaction.
Handling service failure with service recovery is not an easy thing to do, because firms
could never be able to know how dissatisfied their customers are. Moreover, many service firms
and service business researchers (Albrecht et al., 2018; Cheng et al., 2017; Xu et al., 2018)
believe that compensation such as money, coupons, rooms, service, or seats upgrade can offset
customer dissatisfaction. However, Krishna (2014) found out that although monetary and

312 | I C B 2 0 2 1

View publication stats