Elsevier

Applied Soft Computing

Volume 60, November 2017, Pages 763-774
Applied Soft Computing

Using self-organizing maps to model turnover of sales agents in a call center

https://doi.org/10.1016/j.asoc.2017.03.011Get rights and content

Highlights

  • We predict employee turnover using supervised self-organizing maps (SSOM).

  • We used a combination of performance and personality traits as predictor variables.

  • When comparing performance of the classifier with and without personality traits, we found that personality traits help to improve significantly the performance of the classifier.

  • Using capabilities of SSOM, we showed that it is possible to find some personality profiles of agents that make them more or less susceptible to withdraw or to stay in the organization.

Abstract

This paper proposes an approach for modeling employee turnover in a call center using the versatility of supervised self-organizing maps. Two main distinct problems exist for the modeling employee turnover: first, to predict the employee turnover at a given point in the sales agent's trial period, and second to analyze the turnover behavior under different performance scenarios by using psychometric information about the sales agents. Identifying subjects susceptible to not performing well early on, or identifying personality traits in an individual that does not fit with the work style is essential to the call center industry, particularly when this industry suffers from high employee turnover rates. Self-organizing maps can model non-linear relations between different attributes and ultimately find conditions between an individual's performance and personality attributes that make him more predisposed to not remain long in an organization. Unlike other models that only consider performance attributes, this work successfully uses psychometric information that describes a sales agent's personality, which enables a better performance in predicting turnover and analyzing potential personality profiles that can identify agents with better prospects of a successful career in a call center. The application of our model is illustrated and real data are analyzed from an outbound call center.

Introduction

Reducing the employee turnover rate is a challenge for human resource managers. This is particularly critical in certain types of jobs characterized by high workload levels, subject to constant supervision, requiring routine and monotonous tasks that lead to employee fatigue and frustration. One example of this is the call center industry. This paper proposes to use self-organizing maps (SOMs) to model employee turnover in call centers.

The employee turnover rate worldwide in call centers is 20%, which includes promotions, voluntary resignations, retirements and dismissals [1]. However, there is a high degree of variability in this number in emerging and industrialized countries. For example, in emerging countries like India, the turnover rates reach 39%, in Brazil 26% and Poland 25%, whereas industrialized countries like Sweden, Germany and Austria they are 10%, 13% and 4%, respectively [1]. These figures can be even higher when dealing with workers with a short tenure and little work experience. In the previously mentioned emerging countries, the percentage of the workforce with a tenure shorter than one year is about 40%, whereas in industrialized countries it is only 27%. These numbers indicate that a significant proportion of personnel who start working in call centers end up leaving after a year. This supposes high operating costs in recruiting, selecting and training new employees.

We suggest that there are two approaches to reduce turnover: the first it is to change elements of the work environment to eliminate the negative traits that reduce the worker's well-being. The second is to screen out in the recruitment and hiring process, identifying those individuals with a greater likelihood of leaving [2], [3]. This work focuses on the second alternative, using a SOM-based model to predict and analyze the turnover process using a combination of psychometric and performance attributes. Therefore, the contribution of our paper is both methodological and empirical.

Call centers have the technology to record and capture a range of variables that account for a sales agent's activities. This advantage makes it possible to focus efforts on the second alternative. In other words, the availability of data enables a data mining approach to discover patterns in the information, which permits the observation of certain traits in individuals that are more susceptible to leaving or to initiating a turnover process in the early stages of recruitment and hiring. However, the application of machine learning models in this type of problem is limited.

To our knowledge, almost all the studies that have dealt with the issue of turnover with supervised learning algorithms, use sociodemographic attributes or some proxy of the agent's performance. For example, in the problem of predicting performance of sales agents in a call center, it has been shown that sociodemographics attributes do not have predictive power to discriminate between high performers from the ones with low performance [4]. Instead, operational attributes are more useful to predict turnover, because they indicate what an agent does in his labor hours in order to make a sale, and are usually objective information that arises from measurement instruments such as records of hours worked, number of sales, hours talked [5]. The main takeaway from previous works is first, the use of psychometric information as input attributes that feeds a machine learning algorithm, with the purpose of training a classifier that discriminates turnover behaviors, and second to discover combinations of psychological and operational profiles captured from the non-linear dependencies learned by the SOM.

The current study uses a novel dataset that includes psychometric information of the individuals, which allows identifying patterns of permanence or quitting from a call center, given certain levels of psychological factors that describe the employee. To our knowledge, we have not found machine learning applications for this particular problem with this type of data, which is an original contribution of this study. Thus, we demonstrate a successful application of a classifier to predict the turnover in a call center using a combination of each sales agent's performance variables and personality traits from the well-known Big-5 personality scale [6], [7]. To achieve this goal, we proposed to take advantage of a trained Self-Organizing-Map (SOM) to create a classifier which allows us not only to predict the result of a behavior (job turnover), but also to use it in the opposite direction, i.e., to discover under what levels of operational and psychological attributes there is a greater propensity for the employee to stay or leave. From the methodological point of view, this twofold capacity is the greatest novelty and difference in relation to previous works.

There is a wide range of classifiers to predict turnover. Some have already been used successfully (see Section 2.2). If the only aim were to achieve the best predictive performance, classifiers based on support vector machines or neural networks or more complicated schemes could be trained and adjusted to obtain excellent accuracy rates; however, these -black box- classifiers do not permit an understanding of the potential relations between attributes that explain the phenomenon. We claim it is possible to achieve suitable classification performance rates for this data domain, and at the same time greater knowledge of the turnover process using SOMs [8], [9] as a technique of vector quantization that produces an approximation to a continuous probability density function of a vector input (for example, a sales agent's attributes) with information from only two months of work. Aside from the classic advantages of this type of neural network, the SOM has the capacity to map any non-linear relationship without assumptions about the data [10]. In this sense, SOM allows us to estimate the values of an interest variable y (e.g., levels of extroversion of a sales agent) through a simple function that uses the idea that each codevector of each neuron of the SOM, represents a local estimation of the training data.

Psychometric attributes were used in this work that describe a sales agent's personality and attributes of his performance in order to: (1) predict turnover of a particular agent in a specific period and (2) determine the necessary conditions in a sales agent so he stays with (or leaves) the organization. In order to fulfill these two objectives, we conducted a series of simulations with real performance and sales agents personality data from an outbound call center. Then, supervised versions of SOMs were trained as classification models, demonstrating that the combination of personality attributes improves the performance of the classifier. Finally, another series of simulations with trained supervised SOMs was conducted to discover the relations between personality traits and turnover levels to differentiate what levels of personality traits make an agent more prone to stay in a call center from those with a lower survival in the organization.

The contribution of this paper is two-fold: (1) For the first time it has been possible to incorporate psychometric information to a model of machine learning in order to establish a classification model that discriminates individuals that will remain from those that do not, in the company at some moment in time. This information, along with operational sales agent records, allows us to obtain an acceptable classifier performance. (2) To find relationships between variables that determine or help to explain the turnover through a class-change model of an instance. We are not only interested in getting the best classifier for predicting who leaves the company or stays in it, but also understand the levels of certain attributes that make sales agents more or less susceptible to stay or resign. For this purpose, the SOMs are exceptionally versatile, unlike other black-box models [11] used in other studies that can achieve exceptional performance, but they do not offer the opportunity to acquire interpretability and knowledge of what are the interactions between variables that must occur so that, for example, an abandonment of a sales agent is triggered.

This paper is organized as follows: In Section 2, the context of the turnover problem in call centers is explained and a review is made of machine learning studies that examine this issue. We briefly introduce the unsupervised SOM and XY-fused network in Section 3. In this section we also explain our proposal to determine non-linear dependencies among unknown attributes of an input vector versus different levels of well-known attributes in order to better understand the process that triggers a change of state. In Section 4 a case study is introduced as an example of the application of the SOM-based model to predict turnover in a call center. Section 5 shows the results of the classification activities and the turnover analysis, and finally Section 6 gives conclusions.

Section snippets

Context

There are various explanations for what triggers an employee's intention to leave. In the particular case of the call center industry, a work design for mass production services where the goal is to maximize volume and minimize costs, the work has been standardized and automated to a level where the employees’ abilities, discretion and time cycles are minimal [12]. This Taylorian approach to work assumes the workers are replaceable parts. And indeed, the high turnover levels in this industry

Unsupervised SOM

The SOM is an unsupervised learning algorithm [29], consisting of a network of units or interconnected neurons that are distributed spatially in such a way as to preserve a given topology. Usually, this topology is represented on a bi-dimensional plane using a hexagonal or rectangular grid. Before starting to train the network, the units must be initialized, usually assigning vectors from the training set randomly. A competitive learning algorithm is used to train the network. The idea of this

The case study

The data used in this study contains operational information on each of the sales agents in an outbound call center during the months they worked in the organization. The call center is mainly dedicated to telemarketing and selling fire, theft, accident and supplementary health insurance as well as cell phone plans. There is also information measuring personality traits when they applied for a job in the organization. We combined these two sources of information: operational performance and

Topology of the network and learning configuration

There is no explicit rule that can define the number of neurons, the length or width of the map. One heuristic consists of using the squared root of the instance number multiplied by five [47]. For the test sample size (10% of the total sample), this is 101 units. Different simulations with different numbers of units (between 30 and 100) did not reveal any significant changes in the performance of the classifier. Finally, a total of 36 neurons was taken as an option, with a 6×6 configuration,

Conclusion

In this work we have shown the capacity of the SOM not only to discover non-linear patterns in the set of operational and personality attributes of sales agents in a call center, but also as a means of classification among agents who manage to remain in the organization after a test period from those who do not manage to remain for this time. Additionally, we showed that the trained X-Map and Y-Map from the XY-Fused Network can be used to predict any set of attributes with no commitments

Acknowledgments

The authors would thank Alejandro Ventura, CEO of the Call Center Hispanicglobal for their willingness and support during the implementation of the model. This work was partially supported by Fondecyt Research Scholarship (Chile), grants No: 11160072, Basal (CONICYT)-CMM (G.A.R) and the Research Center Millennium Nucleus Models of Crisis (NS130017) (G.A.R).

References (52)

  • M. Sokolova et al.

    A systematic analysis of performance measures for classification tasks

    Inf. Process. Manage.

    (2009)
  • D. Holman et al.

    The global call center report: International perspectives on management and employment, Tech. rep.

    (2007)
  • A. Adorno

    Understanding and combating call center employee turnover

    Custom. Contact World Mag.

    (2000)
  • N. Petouhoff

    Recruiting and Retaining Call Center Employees

    (2006)
  • M. Valle et al.

    Turnover prediction in a call center: behavioral evidence of loss aversion using random forest and naïve bayes algorithms

    Appl. Artif. Intell.

    (2015)
  • R. McRae et al.

    An introduction to the five-factor model and its applications

    J. Personal.

    (1992)
  • M. Barrick et al.

    The big five personality dimensions and job performance: a meta-analysis

    Pers. Psychol.

    (1991)
  • T. Kohonen

    Self-organized formation of topologically correct feature maps

    Biol. Cybern.

    (1982)
  • S. Haykin

    Neural Networks: A Comprehensive Foundation

    (1999)
  • P. Cortez et al.

    Opening black box data mining models using sensitivity analysis

  • R. Batt et al.

    The viability of alternative call centre production models

    Hum. Resour. Manage. J.

    (2002)
  • J. Singh

    Performance productivity and quality of frontline employees in service organizations

    J. Mark.

    (2000)
  • S. Deery et al.

    Work relationships in telephone call centers: understanding emotional exhaustion and employee withdrawal

    J. Manage. Stud.

    (2002)
  • D. Holman

    Employee Stress in Call Centres. Call Centres and Beyond: The Human Resource Management Implications

    (2001)
  • D. Holman et al.

    The effects of performance monitoring on emotional labour and well-being in call centers

    Motiv. Emotion

    (2002)
  • M. McCulloch et al.

    Using person-organization fit to select employees for high-turnover jobs

    Int. J. Select. Assess.

    (2007)
  • Cited by (0)

    View full text