Elsevier

Knowledge-Based Systems

Volume 28, April 2012, Pages 97-104
Knowledge-Based Systems

Predicting customer churn through interpersonal influence

https://doi.org/10.1016/j.knosys.2011.12.005Get rights and content

Abstract

Preventing customer churn is an important task for many enterprises and requires customer churn prediction. This paper investigates the effects of interpersonal influence on the accuracy of customer churn predictions and proposes a novel prediction model that is based on interpersonal influence and that combines the propagation process and customers’ personalized characters. Our contributions include the following: (1) the effects of interpersonal influence on prediction accuracy are evaluated while including determinants that other researchers proved effective, and several models are constructed based on machine learning and statistical methods and compared, assuring the validity of the evaluation; and (2) a novel prediction model based on interpersonal influence and information propagation is proposed. The dataset used in the empirical study was obtained from a leading mobile telecommunication service provider and contains the traditional and network attributes of over one million customers. The empirical results show that traditional classification models that incorporate interpersonal influence can greatly improve prediction accuracy, and our proposed prediction model outperforms the traditional models.

Introduction

Preventing customer churn is an important task for many enterprises, especially in matured industries, including telecommunications [15], [22], [1] and finances [42]. Achieving it requires churn prediction, which is defined as identifying customers who tend to switch to other service providers. Findings from previous research can be categorized according to two aspects. First, churn determinants are analyzed and verified using customer behaviors in various industries. Some attributes, including customer satisfaction, switching costs, customer demographics, tendency to change behavior, and service usage, have been found to be common churn determinants [15], [22], [1]. Second, researchers have proposed prediction models based on machine learning methods, including the decision tree, neural network, and support vector machine [18], [41], [8], [7], or statistical methods, including logistic regression, survival analysis, and Markov chain [21], [24].

Though many practices benefit from these valuable results, one limitation exists. Many researchers have assumed implicitly or explicitly that a customer’s decision to switch service providers is independent of other customers’ decisions. Most prior research focuses exclusively on individual customers, without accounting for any interpersonal influence [9], as they measure each customer’s perception and behavior independently. Typically, many explanatory variables are collected on each customer and used in multivariate prediction modeling. In reality, customers’ behaviors not only depend on their own perceptions and subjective desires but also interplay with each other. Thus, customers’ choices are interdependent [45].

This paper investigates the effects of interpersonal influence on the accuracy of predicting customer churn and proposes a novel prediction model based on interpersonal influence that combines the propagation process and customers’ personalized characters. Our contributions include the following: (1) the effects of interpersonal influence on prediction accuracy are evaluated while including traditional attributes (i.e., customers’ personalized characters) that other researchers proved to be effective, and several models are constructed based on machine learning and statistical methods and compared, assuring the validity of the evaluation; and (2) a novel prediction model based on interpersonal influence and information propagation is proposed. The dataset used in the empirical study was obtained from a leading mobile telecommunication service provider and contains the traditional and network attributes of over one million customers. The empirical results show that traditional prediction models incorporating interpersonal influence can greatly improve prediction accuracy, and our proposed prediction model outperforms the traditional models.

This paper is organized as follows. Section 2 discusses several related works. Section 3 provides the churn determinants used in our models. In Section 4, several classification models are separately constructed based on different attributes and methods, and a propagation-based model is proposed. Section 5 discusses the experimental results and their implications. Section 6 gives our conclusions.

Section snippets

Literature

Customer churn prediction can be regarded as a classification problem, in which each customer is classified into one of two classes, churn or non-churn. Machine learning and statistical methods are the most widely used approaches for classification problems. Popular churn prediction methods include logistic regression [33], decision trees [18], neural networks [41], support vector machines [8], and evolutionary algorithms [3]. However, many previous studies solely focused on customers’

Determinants of customer churn

For clarity and comparison, we classify customer churn determinants into two categories: network and traditional attributes. Network attributes measure interpersonal influence and describes each customer’s local topology in customer contact network and his or her relationships with their neighbors. The other attributes are included in the traditional attributes category, which has been frequently discussed in previous research.

All attributes used in our study were obtained from the

Models

In this study, we employ two types of models. The first are classification models, which are built based on machine learning or statistical methods. We use the classification models to examine whether interpersonal influence (i.e., network attributes) can improve prediction accuracy. The second is our proposed propagation model that combines a propagation process and customers’ personalized characters.

Dataset description

In this study, the experimental data were obtained from December 2008 to June 2009 from a leading mobile service provider and comprise more than one million customers. The customers who had a normal status and still used services in March 2009 were defined as the “basis customers” for prediction. Some basis customers churned from April 2009 to June 2009, during which churn status was identified if the customers never used services in the period and their contacts were terminated by the mobile

Conclusion

In this paper, we investigate the effects of interpersonal influence on the accuracy of predicting customer churn. Interpersonal influence is represented by network attributes that refer to interactions among customers and the topologies of their social networks. We compared the prediction results of traditional attributes-based models, network attributes-based models and combined attributes models and found that incorporating network attributes into predicting models can greatly improve

Acknowledgments

We thank the editor and two referees for many helpful comments. This work was partially supported by the Project 70901009 of National Science Foundation for Distinguished Young Scholars, and the Youth Research and Innovation Program 2009RC1027 in Beijing University of Posts and Telecommunications.

References (45)

  • D. Van den Poel et al.

    Customer attrition analysis for financial services using proportional hazard models

    Eur. J. Oper. Res.

    (2004)
  • K.L. Ailawadi et al.

    An empirical analysis of the determinants of retail margins: the role of store-brand share

    J. Marketing

    (2004)
  • W.H. Au et al.

    A novel evolutionary data mining algorithm with applications to churn prediction

    IEEE Trans. Evolut. Comput.

    (2003)
  • R. Bolton et al.

    Implications of loyalty program membership and service experiences for customer retention and value

    J. Acad. Mark. Sci.

    (2000)
  • L. Breiman et al.

    Classification and Regression Trees

    (1984)
  • J.J. Brown et al.

    Social ties and word-of-mouth referral behavior

    J. Consum. Res.

    (1987)
  • K. Dasgupta et al.

    Social ties and their relevance to churn in mobile telecom networks

  • Z. Dezso et al.

    Halting viruses in scale-free networks

    Phys. Rev. E

    (2002)
  • P. Domingos

    Mining social networks for viral marketing

    IEEE Intell. Syst.

    (2005)
  • J. Goldenberg et al.

    The role of hubs in the adoption process

    J. Marketing

    (2009)
  • M. Granovetter

    Economic action and social structure: the problem of embeddeness

    Am. J. Sociol.

    (1985)
  • A. Gustafsson et al.

    The effects of customer satisfaction, relationship commitment dimensions, and triggers on customer retention

    J. Marketing

    (2005)
  • Cited by (55)

    • tcc2vec: RFM-informed representation learning on call graphs for churn prediction

      2021, Information Sciences
      Citation Excerpt :

      In addition, applying different transformations (e.g. average [60]; ratio/percentage [11,29,45]) is quite common. Several studies also consider slicing according to the churner related information, which can again be considered as one of the dimensions (e.g. total interaction frequency with churners [11,29], and total call volume to/from churner neighbors (seconds) [11,60]). We refer the interested reader to a more comprehensive tabular overview of different variations concerning detailed RFM in our previous study [36].

    • Churn modeling with probabilistic meta paths-based representation learning

      2020, Information Processing and Management
      Citation Excerpt :

      Despite the presence of a variety of RFM variables in the literature, to this end, only three different versions of RFM variables are considered: (1) Summary-RFM (denoted by RFMs) whereby RFM variables are calculated considering all customer interactions; (2) Detailed-RFM (denoted by RFMd), whereby customer interactions are divided according to call direction and destination into three subcategories: (a) outgoing towards home network, (b) outgoing towards other networks and c) incoming and each of R, F, M is calculated according to these; and (3) Churn-RFM (denoted by RFMch) whereby RFM variables are calculated only with respect to interactions with customers who are known to have churned. The latter is motivated by the fact that previous works on churn prediction in telco have recognized the importance of connections with already churned customers (Dasgupta et al., 2008; Modani, Dey, Gupta, & Godbole, 2013; Zhang, Zhu, Xu, & Wan, 2012). In this section, we first explain the construction of different types of guided random walks, through the concept of meta path (types).

    • On the operational efficiency of different feature types for telco Churn prediction

      2018, European Journal of Operational Research
      Citation Excerpt :

      However, none of the previous studies devises a customized way of combining many different feature types in the way our study does (see Table 1). Third, even though some works compare different modeling techniques and feature categories together (Backiel et al., 2014; 2016; Dasgupta et al., 2008; Zhang et al., 2012), comparisons of different modeling techniques are out of the scope of this study. Fourth and most importantly, other studies only consider a PP perspective, except partially for Kim et al. (2014), while our study takes into account the OE/PP trade-off.

    View all citing articles on Scopus
    View full text