The term “machine learning” was coined in 1959 by Arthur Samuel (1959, p. 1). Over the subsequent decades, there has been little consensus on the exact definition of machine learning, as Table 1 indicates.
Table 1. Machine learning: Definitions
Definition References
“[A] field ... concerned with the question of how to construct computer programs that automatically improve with experience.”
Mitchell (1997), p. xv
“[P]rogramming computers to optimize a performance criterion using example data or past experience.” other kinds of decision making under uncertainty.”
Murphy (2012), p. 1
“[C]omputational models using experience to improve performance or to make accurate predictions.”
Mohri et al.
(2012), p. 1
“[A] field of computer science involving creating and continuously improving algorithms that automatically analyze data to identify patterns or predict outcomes.”
Gualtieri (2015), p. 2
“[T]he science of getting computers to learn and act like humans do, and improve their learning over time in autonomous fashion, by feeding them data and information in the form of observations and real-world interactions.”
Faggella (2016)
The present research adopts a combination of these definitions’ three key commonalities: i.e., it holds that machine learning (1) automatically detects patterns in data, (2) automatically improves with experience, and (3) is used to uncover hidden patterns or predict future outcomes.
2.2. Machine learning, data analytics, and customer relationship management
‧
Machine learning has arisen from the maturing field of AI. As noted in the previous section, it has created a set of useful techniques for automating the tedious but crucial task of discovering patterns in databases. Business analysts’ traditional statistical modeling methods for finding patterns in business data have the disadvantage of requiring that their datasets conform to rigid distribution criteria. Traditional OLAP, meanwhile, requires the user to define the query objectives based on his or her personal knowledge and experience, i.e., to pre-judge the possible dimensions or features of the data that will yield the most valuable information, trends and patterns. As datasets grow ever larger, this already subjective process becomes ever more ineffectual as a means of establishing the causes of business problems (Bose & Mahapatra, 2001; Malik, 2013; Olafsson et al., 2008). In contrast, data-mining techniques base their observations on the data itself, impose fewer restrictions, and produce patterns that are easy to understand. Empirical evidence from various fields indicates that data-mining tools clearly outperform conventional tools such as logit and probit models and LDA (Bart & Dirk, 2005; Bhattacharyya et al., 2011; Jones et al., 2015).
There are three main reasons for data mining’s superiority over traditional data-analytic tools when used in CRM. The first is that it reduces the effort involved in feature selection, because it is able to handle many features simultaneously. In this respect, it contrasts sharply with conventional tools such as classical statistical analysis, which are usually one-dimensional and therefore cannot handle huge feature spaces, leading to classification problems. Because data mining enables business users to distinguish sequences of data features according to their influences on the researched problem, it can help them to better allocate resources to the most profitable group of customers. As suggested by Liu and Yu (2005), CRM “is characterized by many input variables, often hundreds, with each one containing only a small amount of information”: a situation in which “human expert feature selection is known to be suboptimal” (Prinzie & Van, 2008, p. 2). Moreover, conventional statistical tools’ lack of efficient feature-selection algorithms not only makes them time consuming to use, but compounds the seriousness of the problems they may cause when applied in CRM (Bhattacharyya et al., 2011).
The second key advantage of data mining is its ability to handle noisy data, e.g., outliers and missing values. As noted by Prinzie and Van (2008, p. 3), “noise robustness is of the utmost importance to CRM applications typically tormented by noisy data”. Conventional tools like logistic regression have no inherent methods of dealing with missing values and are strongly affected by outliers (Coussement et al., 2017), and therefore are very sensitive to noisy data and the shape or structure of input variables. Data-mining techniques are robust in the face of noisy data, with classification trees (decision tree and random forest) being considered especially resistant to noise and overfitting because of their pruning strategies and ensemble of classifiers (e.g., bagging and boosting). As such, data-mining techniques are likely to require only minimal data intervention by the researcher, and provide better
‧
predictive performance (Bhattacharyya et al., 2011; Jones et al., 2015, 2017; Kotsiantis, 2007;
Prinzie & Van, 2008).
Thirdly, data mining provides business users with the flexibility to model nonlinear relationships among variables. Most data-mining techniques, including neural networks, support vector machines (SVMs), generalized boosting, and random forest, are high-dimensional nonlinear models that represent significant improvements over traditional statistical methods. Simple linear classifiers such as standard logit/probit models and LDA have limited capacity to model nonlinearity and unobserved heterogeneity in the dataset (Jones et al., 2015). Since traditional statistical models deal with nonlinear relationships by assuming f; y = f(x), and by replacing the linear model with a high-degree polynomial function, the form of the nonlinearity must be known from the outset (Jones et al., 2017;
West et al., 1997). The usefulness of classification trees as non-parametric classifiers when given non-linearly associated variables has been widely confirmed. A strength of such trees lies in their ability to deal with heterogeneity, “as separate models are automatically fit to subsets of data defined by early splits in the tree” (Prinzie & Van, 2008, p. 2). Therefore, as suggested by Jones et al. (2015), the newer classifiers such as generalized boosting, AdaBoost and random forest can strongly outperform all other classifiers, and have proved remarkably robust in the face of a wide range data structures and assumptions. Using classifiers that are flexible enough to model nonlinear input-output relationships can provide better overall predictive performance (Jones et al., 2015).
Despite these three distinct advantages of machine-learning techniques, some studies have highlighted their drawbacks. One notable argument against machine learning is its lack of interpretability, or the so-called “black box” problem. Though decision trees are relatively easy to explain, other machine-learning techniques including SVMs and neural networks do not allow their users to decipher the role or behavioral influences of input variables (Jones et al., 2017), rendering such techniques’ final classification logic non-understandable (Bose &
Mahapatra, 2001). As a separate issue, applying complicated machine-learning techniques requires parameter tuning and calibration not only before, but also during the process (Coussement et al., 2017). And lastly, it has been argued that if data is collected properly and prepared well, pure statistical modeling can achieve comparable performance to advanced machine-learning techniques (Bhattacharyya et al., 2011; Coussement et al., 2017; Jones et al., 2015); and in many real-world cases, the comprehensibility and justifiability of the model are of much greater importance to its users than obtaining the highest possible level of predictive accuracy (Martens et al., 2011).
2.3. Data analysis for CRM
‧
increasingly important to managers. As data-mining techniques mature and become more affordable, businesses have shown themselves more willing to adopt these tools, with the wider aim of increasing the effectiveness of their CRM. Swift (2001, p. 12) defined CRM as an “enterprise approach to understanding and influencing customer behavior through meaningful communications in order to improve customer acquisition, customer retention, customer loyalty, and customer profitability”. And Ngai et al. (2009, p. 3) defined CRM as“helping organizations to better discriminate and more effectively allocate resources to the most profitable group of customers”. Some theorists including Swift (2001, p. 12), Parvatiyar and Sheth (2001, p. 5), Kracklauer et al. (2004, p. 4) and Ngai et al. (2009, p. 2) have stated that CRM comprises four-part cycle – customer identification, customer attraction, customer retention, and customer development.
The first phase of the CRM cycle, customer identification, involves analysis of customers’ characteristics, so that the most profitable customers and new/potential customers can be segmented and targeted (Ngai et al., 2009). In the second phase, customer attraction, businesses focus and apply resources to directly targeting those customer segments identified in the first phase, with the aim of transforming them into customers (Ngai et al., 2009). The core of CRM is its third phase, customer retention, which aims to maintain and increase customer satisfaction and prevent customer churn via individual marketing, loyalty programs, and complaints management (Ngai et al., 2009). Lastly, customer development involves customer lifetime value analysis, up/cross selling, and market basket analysis, as businesses strive to maximize individual customer profitability, transaction intensity and customer lifetime value (Ngai et al., 2009).
Broken down into these dimensions, the benefits of data analysis to CRM are as follows:
(1) Customer identification: Enables targeting of the right customer segments and/or detection of the customers who are likely to churn.
(2) Customer attraction: Establishes the best communication channels/processes to use with specific customers.
(3) Customer retention: Helps improve customer satisfaction and the maintenance of long-term customer relationships.
(4) Customer development: Increases customer profitability and customer lifetime value.
Iriana and Buttle (2007) and Krishna and Vadlamani (2016) have divided CRM into three primary types: operational, analytical, and collaborative. Analytical CRM involves the use of data-mining techniques, as it aims to improve companies’ business decisions and market distinctiveness via analysis of customers’ behavioral patterns and other customer data.
Therefore, only analytical CRM will be dealt with in the remainder of this paper, which uses the terms “CRM” and “analytical CRM” interchangeably except where otherwise noted.
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
7
2.4. Critical success factors for CRM implementation
Various studies have discussed the factors critical to the success of CRM implementation (Almotairi, 2008; Alt & Puschmann, 2007; Campbell, 2003; Chen & Chen, 2004; Croteau & Li, 2003; King & Burgess, 2007; Pan et al., 2007). The prior research divides such factors into people, processes, and technology. Regarding the first, all seven of the above-cited studies note that the full support of the top-level management of the organization is essential for effective CRM implementation. Additionally, King and Burgess (2007) and Almotairi (2008) note that explaining CRM strategy to staff can raise their awareness of its objectives and help them change their behaviors. And employees’ incentives to apply CRM in their day-to-day work depends principally on (1) the organization’s evaluation and reward structure, which should be aligned to CRM strategy, and (2) suitable training in how to use the CRM system (Almotairi, 2008; Campbell, 2003; Chen & Chen, 2004; Pan et al., 2007).
Within the category of success-critical CRM processes, scholars have identified two broad categories: (1) knowledge management, i.e., the organization’s knowledge of its customers, markets, competitors, etc., and its capabilities of transferring such knowledge into useful information (Chen & Chen, 2004; Croteau & Li, 2003; King & Burgess, 2007); and (2) organizational redesign, whereby the organization is reorganized based on its CRM strategy, with various functions integrated or otherwise connected to support the flow of customer data (Almotairi, 2008; Alt & Puschmann, 2007; Pan et al., 2007).
The third and final critical CRM success factor, technology or technological readiness, implies not only that the organization’s IT architecture should support its business strategies and business processes in a general sense, but also that – prior to CRM implementation – it should be in a state of total readiness, extending to interface design, data storage capabilities, and system configurations (Almotairi, 2008; Chen & Chen, 2004; Croteau & Li, 2003; King
& Burgess, 2007).
‧
CRM, while focusing on three additional key areas, i.e., the process’ operational, managerial and strategic dimensions. Its unifying question is how change can be introduced into the company’s operations such that machine learning can be merged into its business processes.Action research differs from traditional research in terms of how its subjects and objects are defined (Coghlan & Brannick, 2001). More specifically, action research is “an approach in which the action researcher and a client collaborate in the diagnosis of the problem and in the development of a solution based on the diagnosis” (Bryman & Bell, 2011).
For Coughlan and Coghlan (2002), the four broad characteristics of action research are that it is (1) research in action, rather than research about action; (2) participative; (3) concurrent with action; and (4) a sequence of events as well as an approach to problem-solving. Action research was selected for the present case because the members of the target organization participated actively in a cyclical process of interactive actions and collaborations.
3.2. Research process
As the originator of action research, Lewin (1946, p. 5) conceptualized its implementation as “a spiral of steps, each of which is composed of a circle of planning, action and fact-finding about the result of the action”. Coughlan and Coghlan (2002) also proposed a framework for the action-research cycle, comprising a pre-step – aimed at establishing the rationale for both action and research – and six main steps (Fig. 1; see also Maestrini et al., 2016). The six main steps are (1) data gathering through interviews with organizational members, collection of internal documents, and observation; (2) data feedback, in which the researcher provides the information collected in step 1 to the studied organization’s managers for their analysis; (3) data analysis, which is performed collaboratively by the researcher and managers, thus crucially facilitating the effective implementation of an action or actions; (4) action planning, in which activities are scheduled, and roles and responsibilities assigned to organizational members; (5) implementation, in which the organization executes the planned action(s) with the researcher’s support; and (6) evaluation: reflection on the outcomes of the implemented action(s), so that the next cycle may benefit from the cycle that has just been completed; this is the key to continuous learning.
Each action-research cycle leads to another cycle, and thus continuous planning, implementation and evaluation occur over time; and monitoring can be described as a