Data mining techniques for analyzing bank customers: A survey

Abstract

In today’s business world, identifying the customers and analysis of their behavior is important for banking industry. Customer Relationship Management (CRM) is the process of maintaining profitable customer relationships by delivering customer value and loyalty. Moreover, CRM helps to improve the business relationships with customers. The goal of CRM is to maximize the lifetime value of a customer to an organization. Customer Lifetime Value (CLV) can rank and classify customers based on their lifetime value to identify valuable customers and retain them. There are several models for CLV estimation using the past data of customers. This subject helps organizations in their attempts to retain valuable customers. The banks must use appropriate data mining techniques to extract pattern and information from the existing data to gain competitive advantage. Therefore, data mining techniques have an important role to extract the hidden knowledge and information. The goal of this study is to review data mining techniques used for analyzing bank customers in order to help the banks to better identify their customers and design more efficient marketing strategies. The literature covered in this paper is related to the past seventeen years (2001–2017) and these approaches are compared in terms of data sets, prediction accuracy, and so on. We also provide a list of data sets available for the scientific community to conduct research in this field. Finally, open issues and future works in each of these items are presented.

Keywords

Customer relationship management customer lifetime value bank industry data mining loyalty

1. Introduction

In today’s business environment, it is essential to identify and analyze the customers’ needs to gain a competitive advantage. Managers should try to retain customers and focus on key customers in order to increase their costs and gain profit. Today, customers share their negative experiences together using communication technologies, which can lead to the loss of current customers’ trust. Organizations can identify customers and meet their needs to increase customer loyalty. Therefore, retaining key customers for banks is more beneficial than attracting new customers [1]. In today’s continuously changing competitive business environment, the organizations have to analyze and understand customer needs and behavior. “Customer dynamics” is one of the most important issues to be considered when analyzing customer behavior, because customer behavior is often complex and uncertain in today’s dynamic situation. Considering the dynamics of large organizations can cause improvements [2] and organizations with an understanding of customers’ behavior can improve their marketing strategy using behavioral scoring models that help to analyze the behavior of customers [3]. Obtaining customer satisfaction is a modern approach for quality control in organizations. To reinforce customer orientation, many organizations choose customer satisfaction as their main performance indicator, although it is almost impossible to achieve [6]. The loss of customers in an enterprise can reduce its profits. The cost of attracting a new customer is five times higher than that of retaining the existing subscribers. Therefore, customer retention is a core Customer Relationship Management (CRM) issue. CRM creates a strong link between the organization and customers, which finally increases customer loyalty, and enhances return of investment. Organizations can boost customer relationship management and prevent the loss of key customers [4, 57], an improvement that leads to competitive advantages. Companies recognize that CRM is a fundamental tool for building customer value and this topic helps to increase enterprise value [5, 54]. Organizations try to increase Customer Lifetime Value (CLV) through CRM. There are many methods for evaluating the CLV that can be used by organizations to gain profit. The spectrum of research goals is shown in Fig. 1.

Figure 1.

Research’ goals spectrum.

Figure 2.

Timeline of the researches in the literature.

In Fig. 2, a timeline of researches in the literature is shown. For this propose, this paper reviews the researches belonging to CRM and CLV. The rest of the paper is organized as follows: in Section 2, CRM, CLV, and customer segmentation are briefly introduced. In Section 3, the data sets used in recent researches are reviewed. Section 4.1 introduces data mining (DM) techniques applied on CRM, CLV, and customer segmentation and in Section 4.2, the literature is categorized according to the techniques. We will analyze these researches from different aspects and try to provide a comprehensive overview of the literature on this subject, which can be helpful for researchers interested in this field. Evaluation results of the studied papers are illustrated in Section 4.3. In Section 5, we will pay attention to open issues in this context. Finally, we will comprehensively discuss the researches and their results in Appendix II. This information is collected in one table and can thus be easily compared.

2. Literature review

2.1 CRM and e-CRM

Customer Relationship Management (CRM) is a strategy that allows a business to manage its customer relationships. CRM is known as a process for improving the relationship between organizations and customers. It is a practice for creating tolerable connections between an organization and its customers. The goal of CRM is to forge closer and deeper relationships with customers. The most important factors that affect customer loyalty are satisfaction, trust, commitment, perceived value, perceived quality, intuitive image, empathy and switching barriers. Through management of relationships, CRM attempts to attract and retain customers, increase customer loyalty, customer satisfaction, and creation value for customers. According to studies, the cost of attracting a new customer is about five times more than retaining an existing one. Therefore, enterprises found that it is critical to develop a long-term relationship between themselves and their customers to achieve profitability and customer satisfaction in the long run [7, 10, 37, 38, 50, 51, 52]. CRM includes four dimensions: customer identification, customer attraction, customer retention, and customer development. For example, customer satisfaction has many benefits for firms, such as enhancing firm reputation, attracting and retaining customers, increasing customer loyalty, etc. [8, 13, 15]. Note that in [9], the authors surveyed the existing techniques in the field of CRM and explored some of the main challenges. However, this review did not focus on banking industry and had a general scope. In fact, performance evaluation and CRM in the banking industry are more complicated and involve several factors [49].

The use of Internet for commerce presents an opportunity for businesses to use it as a platform for the delivery of CRM functions on the Web (e-CRM). E-CRM expands the traditional CRM techniques by integrating new electronic channels such as web, wireless, and voice technologies, combining them with e-business applications into an overall enterprise CRM strategy. E-CRM is the adaptation of CRM in an e-commerce environment and helps to create and maintain customer relationship using the net [11, 12].

2.2 CLV

Customer Lifetime Value (CLV) is a pattern of analytical CRM. The goal of CRM is to maximize the lifetime value of a customer for an organization. Evaluation of CLV is important for enterprises in terms of marketing to have an effective CRM. Studies have shown that the past behavior of a customer may not always predict their future behavior. Therefore, we need a metric to predict the future profitability of a customer. CLV is a metric that represents the total net profit of a company from any customer. CLV is a concept in CRM domain for evaluating the customer’s value. In addition, it is an important metric for determining how much money a company intends to spend to acquire new customers and how much repeated business a company can expect from certain consumers. CLV can help you understand your most valuable user segments based on customer value over time and is thus considered as an important subject. CLV can rank and classify customers based on their lifetime value for identifying valuable customers and retaining them, as well as identifying and comparing market segments [7, 13, 17, 18, 58, 59]. There are several models to estimate CLV using the past data of customers through which we can determine the customers who are more profitable than others. This topic helps the firms to retain their valuable customers [16]. CLV models are used to evaluate customer loyalty, recommend right products to right customers, and provide individual marketing decisions for each customer. Some CLV models are based on past behavioral models and some on future customer revenue. In addition, a number of CLV models can prospect the future monetary value of customers [14].

In the following, a summary of the some key CLV models are provided, which are divided into tree corresponding categories [17, 18]:

Table 1
Research’ goals spectrum

Reference	Number of	Number of	Metrics	Data source
	customers/ transactions	clusters
[1]	120	2	• Length of the relationship • Recency • Frequency • Monetary	The data from 120 customers were collected and normalized.
[22]	12,429	5	• CustomerID, • SellerID, • Request date, • ComID (commodities ID) • Box fields	The data include 500,220 sales records in sale of 179 different commodities to 12,429 customers.
[19]	250	5	• Gender • Age • Monthly income • Account type • Account closure history • Account currency • Location	The data set contains information about different characteristics of 250 loan applicants of a commercial bank located in Isfahan Province of Iran. The data were collected during a one-month period.
[7]	3,091	9	• Recency • Frequency • Monetary	The data, including date of last purchase, the frequency and money spent in purchases, are extracted from the customers’ transaction database, from which 3,091 records remained after excluding incomplete data.
[2]	Step 1: 165,800 Step 2: 20000	3	Step1: • Recency • Frequency • Monetary Step2: • T • Recency • Frequency • Monetary	The data source has been the database of company.
[29]	1000	4	• Classification – good credit – bad credit • Segmentation – credit amount – credit history – duration in month – other debtors and guarantors – other installment plans – present employment – present residence – property, purpose – savings accounts and bonds – status of existing checking account
[27]	6.2 million	22	• Age • Demographics/ lifestyle data • Product ownership (type and intensity) • Activity level	Data set is based on the analysis of homogeneous groups instead of individual customers.

Table 1, continued
Reference	Number of	Number of	Metrics	Data source
	customers/ transactions	clusters
[10]	16,384	No	• Customer values • Potential value • Customer loyalty	The data of this study consist of 6-month service data of a wireless communication company in Korea. This data set is composed of 200 data fields and 16,384 records of customers.
[61]	10000	No	• Customer satisfaction • Customer loyalty • Product usage	The data set consist of the
monthly data of 10,000 customers of a bank.
[11]	120	No	• Large bank • medium banks	The data from 120 customers of Thai commercial banks.
[35]	578	No	• Satisfaction • Likelihood of switching	The data include 578 existing customers of five major banks in a US state on the east coast.
[53]	45211	17	• Accuracy rate • Sensitivity analysis • Specificity	The data include 17 campaigns of a Portuguese bank conducted between May 2008 and November 2010.
[5]	149 articles	4	• Marketing • Customer service • Understanding customers	The data from four retail trade publications were analyzed over a 5-year period (2000–2005).
[25]	45,211	2	• Classification accuracy • Sensitivity • Specificity	The data include 17 campaigns of a Portuguese bank conducted between May 2008 and November 2010.
[36]	274	No	• Size of banks	The data come from 278 small and large retail banks in Denmark, Finland, Norway and Sweden.
[6]	160	No	• Personnel of the bank • Products • Image of the bank • Service • Access	Input data consist of 160 private customers and 95 companies.
[15]	7,000	4	• Recency • Frequency • Monetary	Data on purchase transactions, approximately 7,000 records, have been collected.
[17]	50,000	6	• Recency • Frequency • Monetary	The data were row data of 50,000 customers’ transactions from early spring 2008 to late summer 2009.
[14]	22086	3	• Recency • Frequency • Monetary	Data set contains 65,535 transaction records with 22,086 customers.
[31]	296	5	• Repayment behavior • Transaction • Frequency • Monetary • Recency	296 customers of the bank in 2009 were used as the data set.
[21]	55,211	No	• Education • Days –segments • Times-Segments • Age range • Sex • Amount of transaction	A table containing effective debit card account information of 55,211 customers until October 2009, and another table storing over 11.3 million individual transaction records for these accounts from March 2007 to October 2009.

Table 1, continued
Reference	Number of customers/ transactions	Number of clusters	Metrics	Data source
			• Transaction type • Terminal type • Occupation • Block code
[16]	214	8	• Recency • Frequency • Monetary	Data set contains 214 RFM data from customers.
[20]	452	7	• Recency • Frequency • Monetary	Data related to 18-month transactions of the short term accounts of 452 customers whose deposit was about 50 million Rails.
[28]	10,281	4	• Recency • Frequency • Monetary	Data collected during 1997 to 1998 in Singapore, to measure customer’s loyalty during these two years.
[18]	298	No	• Relation length • Recency • Frequency • Monetary	The source data is the real transaction data in the bank, which has 298 observed values collected in the file.
[34]	1000	No	• Recency • Frequency • Monetary	The data for this study include a selection of 1000 best customers.
[17]	50,000	9	• Recency • Frequency • Monetary	Two years’ data of transactions, approximately five million records for 50,000 customers, have been collected with simple stochastic sampling method.
[56]	10471	9	• Length • Recency • Frequency • Periodicity	The data contain almost 2 million purchase transactions of 16024 customers during the period between October 1, 2012 and August 31, 2014.
[33]	• SCPOS- all data set 263,082 • Foodmart data set 251,395	No	• Recency • Frequency • Monetary	The final SCPOS-all data set contained 263,082 transactions and 12,505 items while the final Foodmart data set contained 251,395 transactions and 1560 items.

Models for calculation of CLV: This category includes models that are specifically formulated to calculate the CLV.

Models of customer base analysis: The analysis of past behavior of customers to predict their future behavior.

Normative models of CLV: These models have been developed to understand the issues concerning CLV.

2.3 Customer segmentation

Customer segmentation is the process of dividing customers into distinct and meaningful groups of homogeneous customers on the basis of common attributes. Customer segmentation increases customer satisfaction and the company’s expected profit. Customer segmentation is so important for companies to develop effective marketing strategies. The use of various marketing strategies leads to an enhanced value of customers.

Customer segmentation can be done based on various criteria, including degree of customer loyalty, purchase frequency, purchase volume, demographics, etc. The banks can differentiate themselves from their competitors through customer segmentation. Nowadays, customer segmentation has become important in the banking industry, and much research has been done in this respect. On the one hand, customers are demanding quality services from their banks, so that the banks have to keep their customers by customer segmentation to enhance the quality of their services. On the other hand, identifying valuable customers is an important issue and segmentation is an appropriate method to recognize the customers. Cluster analysis is defined as “partitioning data into meaningful subgroups, when the number of subgroups and other information about their composition may be unknown”. There are many models for customer segmentation like SOM (Self-Organizing Map), fuzzy method, RFM (Recency, Frequency and Monetary), combination of some models (RFM, AHP, and K-means algorithm) and so forth [2, 19, 20, 21, 22, 59, 61].

3. Data sets

All the researches about CRM and banking industry have used a data set. They applied several DM techniques to achieve their goals. In this study, data has been collected from banks and some data are taken from other industries. In Table 1, all of these data sets are shown in brief.

Figure 3.

Supervised techniques used in CRM, CLV and bank customer segmentation.

4. Data mining and its techniques on CRM, CLV, and customer segmentation

4.1 Data mining (DM) techniques

DM techniques can help a company to discover and extract useful information from their data warehouses, which cannot be discovered directly. This information can be used for the prediction of their customer behaviors. DM plays an important role for improving customer relationship management. Organizations use DM to improve their competitive advantage and add value to the customer. Therefore, organizations can discover customer needs to increase competitive advantages, respond to the expectations of customers and offer quality service. In recent decades, DM has attracted a lot of attention in data science for diagnosing heart diseases [23, 52], designing software architecture [40, 41, 42, 43], selecting design pattern [44, 62], traffic accident prediction [63], fraud detection [64], ATM (Automated Teller Machine) management [65] and so on, and its techniques have been applied on almost every subject to analyze data and get accurate and reliable results based on the defined goals. Many studies have employed DM to analyze customer data but some have attempted to discover new DM techniques [8, 5, 17]. The important issue is how we can gain better insight and knowledge through using DM techniques to improve CRM and increase CLV and thereby create more profit for customers and company [2]. DM is a technology for extracting information from customer data, which is useful to identify customer demand, effectively promote customer value, and predict their future behavior. DM can also be considered as a process and a technology that uses statistical algorithms to discover hidden patterns and information from the existing data to gain competitive advantage [9, 25, 26]. Organizations have started using DM technologies to gain customer loyalty and increase the contribution of customer value [46].

DM techniques have been used to predict customer behavior and can obtain previously unknown and potentially useful information (including knowledge rules and regularities) by searching through a database. DM is a stage in Knowledge Discovery in Databases (KDD). In today’s competitive business environment, organizations can improve their competitive advantage by using DM. Various applications are used for marketing, finance, and banking. Applications in these domains help to collect and storage a large amount of data. DM gains models from database and discovers patterns and correlations in data and can discover useful customer behavior patterns from large data repositories. In a dynamic business environment, the companies have to analyze and understand customer needs and behavior. Organizations need to have a deeper understanding of customer behavior. Today, customer behavior is often uncertain and changes over time, so organizations can use a number of methods to predict changes in customer behavior and the company needs to focus on mining changes in databases [2, 15, 24, 26]. As we know, retaining a current customer and acquiring a new one is an important matter in banking industry. Therefore, we should use a number of methods for analyzing customer behavior, which are based on bank databases. Analyzing a bank customer is a difficult task because the bank databases are multi-dimensional and are comprised of monthly account records and daily transaction records. In addition, there are a variety of methods to analyze the customer behavior. In this regard, selection of the appropriate method for analysis is important.

In this paper, we review some DM techniques for analysis of bank customer behavior. DM techniques are divided into several groups as follows:

1.
Supervised
2.
Unsupervised
3.
Evolutionary learning
4.
Other

In the following, we discuss these techniques.
4.2 Data mining techniques used in the literature

4.2.1 Supervised

Several supervised techniques are used in knowledge extraction and some of these techniques have been applied on the bank industry, which are shown in Fig. 3.

Decision tree is a supervised classification technique in DM. Decision tree is a predictive model that can be regarded as a tree. This technique represents sets of decisions. The decision tree is most often used for classification of a data set and can produce understandable rules. Specific decision tree techniques such as Classification and Regression Trees (CART), Chi Square Automatic Interaction Detection (CHAID), and Commercial version (C5.0) are highly efficient [10]. In [8], CHAID and C5.0 decision tree techniques are used for classification of customers and help the banking industry to make decisions. They provided a set of rules that can be applied to a new data set to predict which records will have a given outcome. CHAID analysis is an algorithm used for discovering the relationships between variables. CHAID analysis builds a model or tree to help determine how variables best merge to explain the outcome for a given dependent variable. In practice, CHAID is often used in direct marketing to understand how different groups of customers might respond to a campaign based on their characteristics. In [26], authors used CHAID through Applied Matrix for an initial segmentation modeling. Applied Matrix helps companies increase their competitive advantage, the lifetime value of customer, and decrease cost of customer acquisition. In this study, CHAID identified market segments that were formed by interactions among predictors of a chosen criterion variable like customer age. Another technique of decision tree is C5.0, which is a later version of C4.5 and can use continuous data, information theory, and learning method to build a decision tree. C5.0 creates decision tree from a data set by using Gain and Gain Ratio parameters. In [28], authors have used the C5.0 model to produce rules for predicting the level of loyalty based on demographic variables, on the obtained clusters from k-means and two-step algorithms. Moreover, in [4], C5.0 has been used as a classification technique. CART is a classification tree and a form of binary recursive partitioning. In [27], CART analysis has been used to build a regression tree, which uses the customer base into a set of homogeneous sub-groups for clustering. This technique was applied to determine membership of a set of classes as a function of certain predictor variables.

Neural network is an information-processing technique to capture and represent complex data relationships. A neural network acquires knowledge through learning. Neural networks are organized as a layer set of neurons, which include input layers, hidden layers, and output layers, and the computations of the network are performed in the hidden layer. The layers are made up of a number of interconnected nodes. Neural network has been used for creating predictive models such as customer lifetime value. Neural network has a wide range and can be applied to both supervised and unsupervised DM and to solve estimation problems [25, 26]. In [29], neural networks have been used as the classification model and created a credit scoring. The neural networks are built with the data of the existing customers. Then, all these existing customers are evaluated by the model in order to detect their predicted credit status, which can be good or bad.

In [22], K-Nearest Neighbor (KNN) technique is used to classify and identify the goods that are more favorable to customers, which uses a database in which the data points are separated into several classes to predict the classification of a new sample point. In [22], the recommendation systems designed based on a combination of CF (Clustering features) and KNN were used in business.

In [25], Naïve Bayes (NB) classifier is used, which is based on Bayes’ rule as a simple probabilistic classifier. The researcher used Bayes’ rule as the basis for designing learning algorithms. Bayesian classifiers predict the probability that a sample belongs to a particular class. This technique has high accuracy and fastness to train with simple models and is thus used for large databases. According to [19], if the number of clusters is not determined, we can calculate their number by using Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC).

4.2.2 Unsupervised

Unsupervised is a DM technique for extracting hidden structure from unlabeled data, which is shown in Fig. 4. It does not require the humans to have a foreknowledge of the classes, and mainly uses clustering algorithms to classify an image data.

Figure 4.

Unsupervised techniques used in CRM, CLV and bank customer segmentation.

In [2], the GSP (Generalized Sequential Pattern) algorithm is used to extract the sequential patterns. It is a popular algorithm used for sequence mining, which extracts the patterns that appear more frequently than a user-specified minimum support while maintaining their item occurrence order. In [21], Apriori algorithm is used to determine the group of customer by creating customer profiles and finding the relevant clustering rules.

In [3], SOM (Self-Organizing Map) has been used. It is an unsupervised learning technique that relates multi-dimensional data. This technique is an artificial neural network algorithm that contributes to mapping high-dimensional data into a two-dimensionally represented space. SOM has been used in a wide range of applications, including financial data analysis, medical data analysis, time series prediction, and industrial control. It is built with customer’s data, which include variables from account and transaction data sets. In this model, all the customer’s data are used to build the behavioral scoring model in order to predict potential customer behavior. In SOM, it is difficult to predefine a network size without obtaining prior knowledge about the organization of the data to achieve acceptable results. Therefore, the use of a Growing Hierarchical Self-Organizing Map (GHSOM) has been suggested.

Chain model is another unsupervised technique. In [27], a model has been used with a combination of first-order Markov chain modeling and CART. This model is based on the analysis of homogeneous groups instead of individual customers. The chain model has been used in marketing, including customer lifetime valuation.

Figure 5.

Evolutionary techniques used in CRM, CLV and bank customer segmentation.

Clustering is a popular unsupervised learning technique. It is used for finding classes or groups of a data set with most similarities in the same cluster, while the dissimilar objects are in a different cluster. Clustering is a DM technique used to divide data into related groups without a foreknowledge of the group definitions. Clustering techniques are classified as hard clustering and fuzzy clustering. In fuzzy cluster or soft cluster, each point may belong to two or more clusters with different degrees of membership. The well-known hard clustering algorithm (K-means) and Fuzzy clustering algorithm are mostly based on Euclidean distance measure. FCM algorithm is a well-known Fuzzy Clustering method that permits one point to belong to two or more clusters based on fuzzy logic. In [7], FCM algorithm is used to cluster data into nine optimum clusters based on three values of recency, frequency, and monetary. In [1], fuzzy clustering was applied to collect and normalize data from 120 customers based on four different variables, namely length of the relationship, recency of trade, frequency of trade, and monetary value. K-means is an unsupervised DM technique and the most popular hard clustering technique that divides data into groups, and the objects in each cluster are very homogenous and dissimilar with other clusters. Each data point belongs to only one cluster. This technique requires previous knowledge about the number of clusters. It takes the input parameter $k$ (the number of clusters) and partitions a set of n objects into $k$ clusters. In [2], K-means algorithm has been used and $k$ objects have been randomly selected as the initial centers of clusters. Other objects were assigned to the closest clusters based on the distance between the object and the center of cluster. In [15], K-means algorithm was applied for customer segmentation in order to assess CLV for each segment. In [30], K-means algorithm was used as clustering algorithm for clustering the bank’s customers. K-means algorithm was executed eight times and Dunn index was calculated in each time, because this algorithm needed this calculation. The optimum number of clusters ( $k$ ) was found to be four. K-means has been often used as a credit and behavioral ranking model for customers of banks. These models classified customers into different clusters and K-means was applied for ranking new customers in terms of credits using a number of determinants like age, marital status, income, etc. The banks used this ranking for predicting future purchasing behavior of customers or their credit status [31]. In [31], K-means was used to build clusters of customers through RFM variables. To compute K-means, first all $m$ objects were portioned into $K$ initial clusters. Then, the objects whose Euclidean distances were nearest to the mean of cluster were assigned to that cluster. Centroid was recalculated for the clusters with a new item or a lost item. The process was repeated until there would be no more reassignment. In [21], K-means algorithm was used to divide customers based on recency, frequency, and monetary behavioral scoring predictors into three major profitable groups of customers: pamper user, transactor user, and raring user. As already mentioned, calculating bank customer’s CLV is an important matter. Therefore, in [16], the authors used K-means clustering to determine CLV and segment customers based on recency, frequency, and monetary. Moreover, K-means algorithm was applied for cluster analysis of bank data to group customers in [18]. First, raw data were separated into seventeen clusters as favorites; then, the distance of each customer to the center of its cluster and the error function were calculated.

CF tree is a height-balanced tree that stores the clustering features for a hierarchical clustering. Each entry in CF tree represents a cluster of objects. In [19, 22], the authors have used CF tree. For example, in [22], CF is a well-known technique in recommendation systems, which can be categorized as neighborhood-based and model-based techniques. The CF is calculated based on the nearest distinctive neighbor for each cluster of customers.

4.2.3 Evolutionary learning

As shown in Fig. 5, another DM technique is evolutionary learning. This technique can be used in any classification-based prediction scenario. In addition, it helps to predict the value of a user-specified goal attribute based on the values of other attributes. For example, the banks have used this technique to predict credit. Genetic algorithm (GA) is a meta-heuristic algorithm used for data clustering. In [1], this algorithm was used for customer clustering.

4.2.4 Other

In addition to the above-mentioned techniques, some researches have attempted to suggest alternative techniques, which we have categorized as “Other” and are shown in Fig. 6.

Figure 6.

Other techniques used in CRM, CLV and bank customer segmentation.

Delphi method is used in [32] for finding actual CRM definition and customer’s characteristics in the future. This method is based on collective intelligence for finding a common consensus. How a business can gain knowledge through using DM techniques to support intelligent decision making in customer dynamics management is an important subject.

The RFM model is a common well-known method for customer dynamics segmentation. In [15, 47, 48], RFM can be considered as the most powerful and behavior-based model to implement CRM that extracts customer past information by using specific criteria. RFM values are defined as follows:

•

$R$ (Recency): How recently a customer has made a purchase. Customers who have recently purchased from you are more likely to purchase again from you.

•

$F$ (Frequency): How often a customer makes a purchase. The higher the frequency is the chances of customers responding to your offers, which indicates a higher degree of loyalty.

•

$M$ (Monetary): How much money a customer spends on purchases. Customers who have spent higher are more likely to purchase based on the offers, and the company should focus more on such customers.

RFM is a technique for customer behavioral analysis that can effectively investigate customer values. In [24], the RFM scoring model is used to transform the customer behavioral variables. By RFM scoring, customers are segmented into various target markets in terms of customer value. RFM valuables are important hidden variables in the database. In many studies, it has been shown that the higher $R$ and $F$ , the higher will be the probability of new transactions with the customer, and the higher $M$ , the more will be the probability of customer return. Some researchers used this model for ranking customers of Export Development Bank in terms of credit. Therefore, in [31], RFM variables were extracted from the Export Development Bank’s database and were accordingly normalized. Then, the variable weights were calculated using FAHP (Fuzzy AHP), and finally, a value for each customer was estimated. In [2], the RFM model is used to detect customer segment over 8 three-month periods. Moreover, in [3], the RFM values are used to classify three profitable groups of customers as follows: revolver user, transactor user, and convenience user.

LRFM is the extended version of RFM model that is used to take the length of the relationship into account. Authors in [18] represented a model to calculate customer life time value (CLV) based on LRFM customer relationship model, which consists of four dimensions: relation length ( $L$ ), recent transaction time ( $R$ ), buying frequency ( $F$ ), and monetary ( $M$ ) in banking services.

Figure 7.

Percentage of different DM techniques used in CRM, CLV and bank customer segmentation.

Some studies have added the count item to RFM model and implemented the RFMC model. The results revealed that the count item is not so useful and the outcome of RFM model was better than RFMC model [15]. Therefore, CLV is calculated based on weighted RFM method for each segment. In [56], an augmented RFM model called LRFMP to gain deeper and reasonable insight about customers are proposed.

A weighted RFM integrates AHP and DM into customer segmentation. Some researchers proposed WRFM (Weighted RFM) instead of RFM. They dedicated weights to $R, F$ , and $M$ depending on characteristics of the industry. These weights are different based on characteristics of the industry. For example, in some researches, placing the highest weighting on frequency, followed by recency has been suggested, with the lowest weighting on the monetary measure. The AHP method is a hierarchical process that is used to determine the weights of the RFM variables based on expert point of view in sales department. AHP is a well-known tool for decision-making in an analysis. In [20], researchers have proposed a combination of RFM with AHP and K-means algorithm and then segmented it with a group of customers from one of the big national banks in the country.

4.3 Evaluation of research’s techniques

In the previous section, we studied DM technique used by some researchers for analyzing the customer behavior of banks and another finance industry. Now, we are going to evaluate these techniques based on the results of references. In Fig. 7, the usage percentage of different techniques in this literature is shown. As can be seen in this Fig. 7, unsupervised learning techniques have the highest usage percentage in this literature.

4.3.1 Evaluation criteria

There are a number of criteria to evaluate DM techniques used in this study. We define these criteria and will discuss their obtained results. At first, we define TP, TN, FP, and FN according to the recommendation system [45].

•
True Positive (TP): A product was recommended to a customer and the customer bought it.
•
False Positive (FP): A product was recommended to a customer and the customer did not buy it.
•
False Negative (FN): A product was not recommended to a customer and the customer bought it.
•
True Negative (TN): A product was not recommended to a customer and the customer did not buy it.

Accuracy: This criterion is used to evaluate the performance of classification models. Classification accuracy is defined by Eq. (1) [45]:

$\displaystyle\textit{Accuracy}=\frac{\textit{TP}+\textit{TN}}{\textit{TP}+% \textit{TN}+\textit{FN}+\textit{FP}}$ (1)

Recall and Precision: These criteria are used to measure the quality in recommender and information retrieval systems and are defined as follows [45]:

$\displaystyle\textit{Recall}=\frac{\textit{TP}}{\textit{TP}+\textit{FN}}$ (2) $\displaystyle\textit{Precision}=\frac{\textit{TP}}{\textit{TP}+\textit{FP}}$ (3)

Note that the increase in recommendation causes a decrease in Precision and an increase in Recall.

Figure 8.
Results of Dunn index for clustering techniques [30].

F-measure: This criterion leads to a balance in Precision and Recall [45]:

$\displaystyle\textit{F-measure}=\left(2\times\textit{Recall}\times\textit{% Precision}\right)/\left(\textit{Recall}+\textit{Precision}\right)$ (4)

Figure 9.
Comparison the Accuracy of different techniques.

Support: Support is a measure of how frequently any given combination of antecedent and consequent occurs in a data set [45].

Confidence: It is defined by the percentage of cases, in which a consequent appears given that the antecedent has occurred. It essentially measures the strength of an association rule [45].

SSE: In some studies, sum of squared errors (SSE) [45] is investigated to calculate the appropriate number of clusters, which is shown in Eq. (5).

$\displaystyle\textit{SSE}=\sum_{i=1}^{k}\sum_{x\in c_{i}}{{\textit{dist}}^{2}% \left(m_{i},x\right)}$ (5)

Where $k$ is number of clusters, $m_{i}$ is the cluster center, and x is the record belonging to the cluster.

If SSE is closer to 0, it indicates that the model has smaller random error component and is useful for prediction.

RMSE: This is a good measure of accuracy, which compares and defines the forecasting errors of different models by the following equation.

$\displaystyle\textit{RMSE}=\sqrt{\textit{MSE}}$ (6)

MSE: It is the mean square error or the residual mean square [45]:

$\displaystyle\textit{MSE}=\frac{\textit{SSE}}{v}$ (7)

For example, in [1], a combined algorithm (fuzzy c-means cluster and genetic algorithm) was proposed. Therefore, for comparing the efficiency of these algorithms, mean squared error (MSE) was used and its results are shown in Table 2. According to this study, it is clear that the combined clustering algorithm had a lower MSE; therefore, the authors suggested the use of this combination to obtain the most accurate cluster.

Table 2
Timeline of the researches in the literature

Algorithm MSE

Fuzzy c-mean 0.0875

Combination of clustering and genetic algorithm 0.0152

Dunn index: This metric is used for evaluating the clustering techniques. For a given assignment of clusters, a higher Dunn index indicates better clustering. For example, in [30], three techniques (including K-means, Two-Step, and X-means) are used for estimation of customer future value of different customer segments based on the adapted RFM model in retail banking context. In Fig. 8, results of this index are shown.

We studied the usage frequency of each criterion in the literature. Accordingly, the most frequently used criteria are Accuracy, Precision, Recall, and F-measure. As mentioned earlier, Accuracy is the most widely used criterion in all classes of DM techniques, because it provides a good insight for users. Figure 9 shows the accuracy results reported by the researches. This figure gives important information by comparing different techniques and approaches in terms of Accuracy.

According to the results presented in [28], if Two-step and K-means for customer segmentation are used to measure customer loyalty, the K-means algorithm has a poor function in recognizing the medium and very high loyalty levels. The two-step algorithm has a very low Accuracy for recognizing a low level of loyalty.

In [8, 25], some DM techniques were used to classify customers and help the desired organization to make decisions. In [8], CAID and C5.0 decision tree techniques have been applied. According to Fig. 9, it is clear that the applied CHAID technique has more than 82% Accuracy, which is about 74% for C5.0 technique. As shown in Fig. 9, in [25], Neural Network and NB were used to predict the customer behavior and it was concluded that Neural Network is the best model for predictive performance with Accuracy rate of 88.63%.
5. Discussion and conclusion

Algorithm	MSE
Fuzzy c-mean	0.0875
Combination of clustering and genetic algorithm	0.0152

According to our survey, numerous studies have been done on customer behavior until now. Some researchers believe that the importance of customer value in financial services industry is seldom realized. In banking industry, identifying the customers and their needs is an important matter. Assessment of the value of bank customers and determining their impact on the performance of banks are necessary to identify their key characteristics by using customer clustering. The banks can identify their most profitable customers and design marketing strategies for each group of customers by customer clustering. In this literature, we attempted to cover every aspect of this subject and almost reviewed all the studies conducted from 2001 to 2017, which are summarized in Appendix 2. In this paper, we conducted a research about data sets that have been used in recent studies. Moreover, we tried to review and summarize all DM techniques associated with analysis of bank customers. We reported experimental results presented by researches and the criteria used in them. Some studies have used a combination of DM techniques to increase Accuracy of their study. Moreover, there are many subjects in this domain, which needs further studies, and we are going to cite them in the following.

In today’s dynamic environment, identifying customers’ behavior is important for banking industry to gain a competitive advantage. The banks have to analyze their customers to identify their needs, which helps the banking industry to retain valuable customers and increase their CLV. Considering the importance of this subject, numerous studies have been done in this area in recent years. According to these studies, there are some issues that need more attention and we try to mention some of them in the following:

•

Some researchers have used a limited data set; therefore, researchers should prepare a general data set to achieve a more complete analysis and make better decisions in future.

•

Because Accuracy of customer clustering is an important matter, it is necessary that future studies use a combination of fuzzy clustering algorithms and evolutionary algorithms.

•

Numerous studies have been conducted on CLV and the authors of these studies recommend researchers to work on more measures and comprehensively consider them or rather compare various CLV models in a specific industry.

•

Some researchers used WRFM techniques and clustering algorithms based on customer’s value to specify loyal and profitable customers; therefore, we recommend future researches to automatically set the weight of variables.

Footnotes

Appendix

References

Ansari

Riasi

. Customer clustering using a combination of fuzzy c-means and genetic algorithms. International Journal of Business and Management. 2016 Jun 21; 11(7): 59.

Noughabi

Far

Albadvi

. Intelligent Decision Making for Customer Dynamics Management Based on Rule Mining and Contrast Set Mining. InTheoretical Information Reuse and Integration 2016; (pp. 135-155). Springer International Publishing.

Hsieh

. An integrated data mining and behavioral scoring model for analyzing bank customers. Expert Systems with Applications. 2004 Nov 30; 27(4): 623-33.

Chu

Tsai

. Toward a hybrid data mining model for customer retention. Knowledge-Based Systems. 2007 Dec 31; 20(8): 703-18.

Anderson

Jolly

Fairhurst

. Customer relationship management in retailing: A content analysis of retail trade journals. Journal of Retailing and Consumer Services. 2007 Nov 30; 14(6): 394-9.

Mihelis

Grigoroudis

Siskos

Politis

Malandrakis

. Customer satisfaction measurement in the private bank sector. European Journal of Operational Research. 2001 Apr 16; 130(2): 347-60.

Safari

Montazer

. Customer lifetime value determination based on RFM model. Marketing Intelligence & Planning. 2016 Jun 6; 34(4): 446-61.

Jenabi

Mirroshandel

. Using data mining techniques for improving customer relationship management. European Online Journal of Natural and Social Sciences: Proceedings. 2014 Sep 23; 2(3s): pp-3143.

Soltani

Navimipour

. Customer relationship management mechanisms: A systematic review of the state of the art literature and recommendations for future research. Computers in Human Behavior. 2016 Aug 31; 61: 667-88.

10.

Kim

Jung

Suh

Hwang

. Customer segmentation and strategy development based on customer lifetime value: A case study. Expert systems with applications. 2006 Jul 31; 31(1): 101-7.

11.

Sivaraks

Krairit

Tang

. Effects of e-CRM on customer–bank relationship quality and outcomes: The case of Thailand. The Journal of High Technology Management Research. 2011 Dec 31; 22(2): 141-57.

12.

Navimipour

Soltani

. The impact of cost, technology acceptance and employees’ satisfaction on the effectiveness of the electronic customer relationship management systems. Computers in Human Behavior. 2016 Feb 29; 55: 1052-66.

13.

Hosseini

Maleki

Gholamian

. Cluster analysis using data mining approach to develop CRM methodology to assess the customer loyalty. Expert Systems with Applications. 2010 Jul 31; 37(7): 5259-64.

14.

Tabaei

Fathian

. Using Customer lifetime Value Model for Product Recommendation: An Electronic Retailing Case Study. International Journal of e-Education, e-Business, e-Management and e-Learning. 2012 Feb 1; 2(1): 77.

15.

Khajvand

Zolfaghar

Ashoori

Alizadeh

. Estimating customer lifetime value based on RFM analysis of customer purchase behavior: Case study. Procedia Computer Science. 2011 Jan 1; 3: 57-63.

16.

Sohrabi

Khanlari

. Customer lifetime value (CLV) measurement based on RFM model. Iranian Accounting & Auditing Review. 2007; 47: 7-20.

17.

Khajvand

Tarokh

. Analyzing customer segmentation based on customer value components (Case Study: A Private Bank). 2011; 79-93.

18.

Alvandi

Fazli

Abdoli

. K-Mean clustering method for analysis customer lifetime value with LRFM relationship model in banking services. International Research Journal of Applied and Basic Sciences. 2012; 3(11): 2294-302.

19.

Ansari

Riasi

. Taxonomy of Marketing Strategies Using Bank Customers’ Clustering. International Journal of Business and Management. 2016 Jun 21; 11(7): 106.

20.

Rezaeinia

Keramati

Albadvi

. An integrated AHP-RFM method to banking customer segmentation. International Journal of Electronic Customer Relationship Management. 2012 Jan 1; 6(2): 153-68.

21.

Farajian

Mohammadi

. Mining the banking customer behavior using clustering and association rules methods. International Journal of Industrial Engineering. 2010 Dec; 21(4).

22.

Rezaeinia

Rahmani

. Recommender system based on customer segmentation (RSCS). Kybernetes. 2016 Jun 6; 45(6): 946-61.

23.

Homayounfar

Sepehri

Hasheminejad

Ghobakhloo

. Designing a chronological based framework for condition monitoring in heart disease patients-a data mining approach (DM-PTTD). Iranian Journal of Medical Informatics. 2014 Sep 1; 3(3).

24.

Chen

Chiu

Chang

. Mining changes in customer behavior in retail marketing. Expert Systems with Applications. 2005 May 31; 28(4): 773-81.

25.

Bahari

Elayidom

. An efficient CRM-data mining framework for the prediction of customer behaviour. Procedia Computer Science. 2015 Jan 1; 46: 725-31.

26.

Rygielski

Wang

Yen

. Data mining techniques for customer relationship management. Technology in Society. 2002 Nov 30; 24(4): 483-502.

27.

Haenlein

Kaplan

Beeser

. A model to determine customer lifetime value in a retail banking context. European Management Journal. 2007 Jun 30; 25(3): 221-34.

28.

Qiasi

Minaei-Bidgoli

Amooee

. Developing a model for measuring customer’s loyalty and value with RFM technique and clustering algorithms. The Journal of Mathematics and Computer Science. 2012 4(2): 172-81.

29.

Kim

Sohn

. Managing loan customers using misclassification patterns of credit scoring model. Expert Systems with Applications. 2004 May 31; 26(4): 567-73.

30.

Khajvand

Tarokh

. Estimating customer future value of different customer segments based on adapted RFM model in retail banking context. Procedia Computer Science. 2011 Jan 1; 3: 1327-32..

31.

Mohammadi

Bidabad

Nourasteh

Sherafati

. Credit Ranking of Bank Customers (An Integrated Model of RFM, FAHP and K-means). European Online Journal of Natural and Social Sciences. 2014 Jul 1; 3(3): 564.

32.

Triznova

Maťova

Dvoracek

Sadek

. Customer Relationship Management Based on Employees and Corporate Culture. Procedia Economics and Finance. 2015 Jan 1; 26: 953-9.

33.

Yeh

. Discovering valuable frequent patterns based on RFM analysis without customer identification information. Knowledge-Based Systems. 2014 May 31; 61: 76-88.

34.

Carrasco

Blasco

Herrera-Viedma

. A 2-tuple Fuzzy Linguistic RFM Model and Its Implementation. Procedia Computer Science. 2015 Jan 1; 55: 1340-7.

35.

Manrai

. A field study of customers’ switching behavior for bank services. Journal of Retailing and Consumer Services. 2007 May 31; 14(3): 208-15.

36.

Lüneborg

Nielsen

. Customer-focused technology and performance in small and large banks. European Management Journal. 2003 Apr 30; 21(2): 258-69.

37.

Kim

Park

. Integration of firm’s resource and capability to implement enterprise CRM: A case study of a retail bank in Korea. Decision Support Systems. 2010 Jan 31; 48(2): 313-22.

38.

Tohidi

Jabbari

. CRM in Organizational Structure Design. Procedia Technology. 2012 Jan 1; 1: 579-82.

39.

Park

Chang

. Individual and group behavior-based customer profile model for personalized product recommendation. Expert Systems with Applications. 2009 Mar 31; 36(2): 1932-9.

40.

Hasheminejad

Jalili

. SCI-GA: Software Component Identification using Genetic Algorithm. Journal of Object Technology. 2013; 12(2): 3-1.

41.

Hasheminejad

Jalili

. An evolutionary approach to identify logical components. Journal of Systems and Software. 2014 Oct 31; 96: 24-50.

42.

Hasheminejad

Jalili

. CCIC: Clustering analysis classes to identify software components. Information and Software Technology. 2015 Jan 31; 57: 329-51.

43.

Tawosi

Jalili

Hasheminejad

. Automated software design using ant colony optimization with semantic network support. Journal of Systems and Software. 2015 Nov 30; 109: 1-7.

44.

Hasheminejad

Jalili

. Design patterns selection: An automatic two-phase method. Journal of Systems and Software. 2012 Feb 29; 85(2): 408-424.

45.

Alpaydin

. Introduction to machine learning. MIT press; 2014; Aug 22.

46.

Weng

. Revenue prediction by mining frequent itemsets with customer analysis. Engineering Applications of Artificial Intelligence. 2017 Aug 31; 63: 85-97.

47.

Shim

Choi

Suh

. CRM strategies for a small-sized online shopping mall based on association rules and sequential patterns. Expert Systems with Applications. 2012 Jul 31; 39(9): 7736-42.

48.

Dursun

Caber

. Using data mining techniques for profiling profitable hotel customers: An application of RFM analysis. Tourism Management Perspectives. 2016 Apr 30; 18: 153-60.

49.

Gounder

Venkateshwarlu

. Shareholder Value Creation: An Empirical Analysis of Indian Banking Sector. Accounting and Finance Research. 2017 Feb 14; 6(1): 148.

50.

Bahramzadeh

Shokati Moghareb

. Identifying and ranking the factors affecting customer loyalty of private banks in Khouzestan province. Insecond international conference of financial services marketing, Tehran, Iran. Retrieved January 2009; 20: 2016.

51.

Ansari

Riasi

. Modelling and evaluating customer loyalty using neural networks: Evidence from startup insurance companies. Future Business Journal. 2016 Jun 30; 2(1): 15-30.

52.

Rahman

Khan

. An Assessment of Data Mining Based CRM Techniques for Enhancing Profitability. I.J. Education and Management Engineering. 2017; 30-40.

53.

Bahari

Elayidom

. An efficient CRM-data mining framework for the prediction of customer behaviour. Procedia Computer Science. 2015 Jan 1; 46: 725-31.

54.

Fotiadis

Vassiliadis

. Being customer-centric through CRM metrics in the B2B market: the case of maritime shipping. Journal of Business & Industrial Marketing. 2017 Apr 3; 32(3): 347-56.

55.

Michel

Schnakenburg

von Martens

. Effective customer selection for marketing campaigns based on net scores. Journal of Research in Interactive Marketing. 2017 Mar 13; 11(1): 2-15.

56.

Peker

Kocyigit

Eren

. LRFMP model for customer segmentation in the grocery retail industry: a case study. Marketing Intelligence & Planning. 2017 May 6; 35(4): 544-59.

57.

Mitik

Korkmaz

Karagoz

Toroslu

Yucel

. Data Mining Approach for Direct Marketing of Banking Products with Profit/Cost Analysis. The Review of Socionetwork Strategies. 2017; 1-5.

58.

Lycett

Marshan

. Modelling Connected Customer Lifetime Value (CCLV) in the Banking Domain. 2017.

59.

Kahreh

Tive

Babania

Hesan

. Analyzing the applications of customer lifetime value (CLV) based on benefit segmentation for the banking sector. Procedia-Social and Behavioral Sciences. 2014 Jan 8; 109: 590-4.

60.

Estrella-Ramón

Sánchez-Pérez

Swinnen

VanHoof

. A model to improve management of banking customers. Industrial Management & Data Systems. 2017 Mar 13; 117(2): 250-66.

61.

Ekinci

Uray

Ülengin

. A customer lifetime value model for the banking industry: a guide to marketing actions. European Journal of Marketing. 2014 Apr 8; 48(3/4): 761-84.

62.

Chihada

Jalili

Hasheminejad

Zangooei

. Source code and design conformance, design pattern detection from source code by classification approach. Applied Soft Computing. 2015 Jan 31; 26: 357-67.

63.

Hashmienejad

Hasheminejad

. Traffic accident severity prediction using a novel multi-objective genetic algorithm. International Journal of Crashworthiness. 2017 Jan 1: 1-16.

64.

Hasheminejad

Salimi

. FDiBC: a novel fraud detection method in bank club based on sliding time and scores window. Journal of AI and Data Mining. 2018 Mar 1; 6(1): 219-31.

65.

Hasheminejad

Reisjafari

. ATM management prediction using Artificial Intelligence techniques: A survey. Intelligent Decision Technologies. 2017 Jan 1; 11(3): 375-98.

Data mining techniques for analyzing bank customers: A survey

Abstract

Keywords

1. Introduction

2.1 CRM and e-CRM

2.2 CLV

Table 1 Research’ goals spectrum

3. Data sets

4.1 Data mining (DM) techniques

1. Supervised 2. Unsupervised 3. Evolutionary learning 4. Other In the following, we discuss these techniques. 4.2 Data mining techniques used in the literature

4.2.1 Supervised

4.2.2 Unsupervised

4.2.4 Other

4.3.1 Evaluation criteria

Footnotes

Appendix

References

Table 1
Research’ goals spectrum

1.
Supervised
2.
Unsupervised
3.
Evolutionary learning
4.
Other

In the following, we discuss these techniques.
4.2 Data mining techniques used in the literature