Human talent management, selection process, and labor turnover

HRM contributes to achieving the company's strategic objectives by carrying out crucial activities such as fostering favorable relationships between employees, controlling labor turnover, and controlling the internal and external factors that may affect the employee's development (^{Dessler, 2017}). In particular, in low-skilled jobs, several internal and external factors may influence the employee's decision to quit or stay in the job: organizational climate, commitment, satisfaction, flexibility, and company policies (^{Aranibar et al., 2018}). Some other key factors are mentioned, such as staff benefits, working environment, motivation, leadership, career plan, age, place of residence, number of dependents, salary, training, job expectations, the influence of co-workers, and job fit and expectation in several research works (^{Dubey et al., 2016}; ^{Tam & Khoa, 2018}; ^{Al Mamun & Hasan, 2017}; ^{Kangas et al., 2018}). Furthermore, ^{Ranganathan & Kuruvilla (2008)} analyzed three spectra that influence attrition that is related to: i) work environment, ii) employee demographic profile, and iii) psychological factors related to the employee profile.

Regarding Mexico, the manufacturing industry reports that labor turnover is associated with the economy's dynamism, which is correlated directly with the Gross Domestic Product (GDP) and inversely with unemployment (^{Moreno et al., 2015}). During economic expansion, the manufacturing industry generates structural contradictions by demanding job vacancies that are not filled, which weakens the productive process, calling into question the efficiency and effectiveness of the manufacturing industry (^{Herrera et al., 2019}). On the other hand, the Mexican hotel industry reports three leading causes of labor turnover, particularly in Guanajuato state (^{Caldera et al., 2019}): i) compensation systems (salary and non-salary related) are not competitive, ii) personnel recruitment and selection systems are inadequate, and iii) employees are unmotivated due to lack of clarity in their tasks and functions.

^{Mondy & Noe (2005)} suggest that HRM depends on human resources research, which tries to find the answers to the human resources problems that occur and affect the company, including labor turnover. Part of this human resources research focuses on controlling labor withdrawal and turnover rate by trying to understand its roots. Mondy & Noe also attempt to understand the use of technology to improve the HR department's process, for instance, the implementation of data analytics to measure the efficiency of HRM practices and processes and even predict events (e.g., turnover rate and employee behavior and performance).

One of the most important activities of HRM is the recruitment and selection process, which helps to create a pool of talent and potential candidates who might help to reach the business goals (^{Uzair et al., 2017}). A good selection of personnel in a company brings benefits related to the organizational performance of the employees, which leads to achieving organizational objectives. A good selection of personnel in a company brings benefits related to the organizational performance of the employees, which leads to achieving organizational objectives. The recruitment and selection processes are fundamental to selecting the right person for the right job (^{Uzair et al., 2017}; ^{Almeda, 2017}) because these processes shape effectiveness and performance.

Data mining applied in human resource management

The application of data mining techniques in HRM has been reported in related research using data mining as a descriptive trend and data mining as a predictive trend.

The descriptive approach analyzes current HRM data to describe insights and generate knowledge about a particular problem. With this knowledge, action strategies are proposed to solve problems or reduce impacts on the HRM and the company. For example, ^{Gao (2017)} implements data mining methods to address employee turnover problems using the algorithm CART to analyze the most important factors contributing to high employee turnover. The results included a priority table that shows those people who should be retained.

On the other hand, the predictive approach has had a more active role in current research. For example, (^{Sikaroudi et al., 2015}) applied data mining techniques to predict the voluntary turnover of employees in a manufacturing plant in Arak, Iran. This study proposed a decision support system for the human talent department through data mining and evaluating employee characteristics such as age, technical skills, and work experience. In particular, this study attempted to infer the performance of candidates for positions in the company to reduce losses by avoiding hiring unstable and poorly trained candidates.

The research made by ^{Attri (2018)} proposed a model to predict the probability that an employee leaves a position based on previous information from the human talent department. The primary purpose of this model was to find and apply strategies that prevent these behaviors. This proposal is based on the fact that the performance of the employees is predictable, and this knowledge would help to avoid the attrition of valuable employees for a company. ^{Attri (2018)} used criteria given by experts in the area, such as age, sex, department to which the worker belongs, job satisfaction, and education, among many others. It proved that the model could predict employee behavior by 80 %.

Also, ^{Ribes et al. (2017)} implemented common data mining techniques on a data set of 1000 employees and created a turnover prediction model. The resulting model showed that employees low performance, compensation schemes, and business units particularities were factors that predict turnover. The results led to the design of retention policies divided into reassignment programs, rotation of positions, and binding of new employees. Similarly, the studies mentioned above seek to solve the problem of labor turnover through data mining techniques. However, other studies have attempted to apply these techniques to predict employee behavior and performance. For example, the study by ^{Kirimi & Moturi (2016)} proposes applying data mining classification techniques to predict the performance of existing employees in a government enterprise in Nairobi. A job performance classification model was created based on sociodemographic information and previous performance analyzes. This model was used to improve the process of job performance evaluation. As a result, it was obtained that experience was the factor with the most significant positive impact on employee performance, followed by factors such as age, grades, sex, marital status, and performance evaluation score.

As seen in this section, there is still a series of challenges to HR analytic: research efforts still focus on descriptive and predictive analytics (^{Pape, 2016}), but there is an opportunity to implement prescriptive analytics. It is noteworthy that most research is focused on high-skilled employees, disregarding data about low-skilled workforce and outsourcing companies, a sector that is becoming a trend in HRM. In this work, we attempt to provide an analysis of such data by following a data mining approach.

Business understanding

This phase aims to understand the business problem and determine an objective for the data mining project. We focused on analyzing the problems of the HRM in an Outsourcing company.

We conducted a situational analysis in Malumex, a Mexican company in Sonora State. We used the technique of Narrative Research (^{Hernández et al., 1991}), to create a story presented as a case study about the company's growth through the years, the problems presented in HR processes, and the perception of the importance of employee profiles within selection processes and turnover rate monitoring. This technique was selected because it facilitates the use of information provided by people in the company and allows them to have a more secure environment and flexible time in the sessions. Also, we used some confidential company documents, including recruitment, selection, and sales reports and statistics. These documents helped complete the story provided by employees and further understand the company. Furthermore, this helped obtain the context necessary to define a data mining objective.

Meetings and documentary revision led to the creation of the case study. The company's Chief Human Resources Officer (CHRO) and Chief Executive Officer (CEO) were present in the five sessions, which lasted between 40 and 60 minutes. The topics discussed were the company's history, HRM Office's processes, problems related to the employees and turnover rate, strategies implemented to reduce turnover rate, the selection process, the systems and documents used to manage employee data, and consequences of turnover to the company. The meetings were recorded and transcribed to integrate the case study.

Data understanding

The main objective of this phase is to understand the obtained data, its sources, and its characteristics. We collected data from different sources related to employee profiles and recruitment processes of the company, which are associated with turnover according to reported literature (^{Dubey et al., 2016}; ^{Ranganathan & Kuruvilla, 2008}; ^{Sikaroudi et al., 2015}; ^{Attri, 2018}; ^{Kirimi & Moturi, 2016}): i) SuperNOMINA software that included information of employees over the past 20 years, ii) job application forms that included socio-demographic information from the period 2016-2019, and iii) the digital records of job interviews from the period 2017-2019. The three data sources were integrated into a single data set by capturing the information manually in an MS Excel spreadsheet, which resulted in a data set with 20 attributes.

We also made an exploratory analysis of this dataset to understand the types of attributes better. The descriptive analysis included the calculation of measures of central tendency and frequencies using the statistical software SPSS.

Data preparation

This phase aims to transform the obtained data into a suitable format to execute data mining algorithms and create models. One of the biggest challenges in this phase was to create a single data set from the three sources. To achieve this, we extracted a subset from the initial data set that matched the period from the other two data sources. The data set was revised to eliminate information duplicated. Furthermore, we executed five data preparation tasks to transform the variables into a suitable format for the data mining algorithms:

Modeling

The purpose of this phase is to create a model that solves the objective of the project. Once the data set was prepared, we executed two iterations using Rapidminer software: manual and automatic, featuring a selection of variables for the training and testing subsets. The results and performance of the following data mining techniques were compared: Naive Bayes (NB) (^{Chhogyal & Nayak, 2016}), Generalized Linear Model (GLM) (^{Dobson & Barnett, 2008}), Logistic Regression (LR), Random Forest (RF) (^{Couronné et al., 2018}), Deep Learning (DL) (^{Guo et al., 2016}), Decision Tree (DT) (^{Jahan et al., 2018}), and Support Vector Machine (SVM) (^{Van et al., 2016}), as these techniques are the most used for classification problems in related literature. The configuration of the algorithms was set as the default parameters given by the data mining system.

A partition relation of 0.6 and 0.4 was established to create training and testing subsets, respectively. Also, an automatic sampling type with an automatic value of Local Random Seed was established. The training and validation subsets in the second iteration were created following the same rules. An Automatic Feature Selection and Generation were established to create an optimal Feature Set regarding the complexity and error. The feature set was applied to the complete training data to build the final model.

Evaluation

This phase aims to establish the criteria to evaluate the resulting model. The most common performance indicators for data mining models are accuracy and classification error. Other indicators are Area Under the Curve (AUC), sensitivity, f measure, and confusion matrix. For this project, accuracy, classification error, and AUC were observed to determine the performance of the models. An accuracy closest to 1 means that the model has a good classification performance, while a classification error closest to 0 means the possibility of error is the lowest in the model. According to ^{Narkhede (2018)}, the AUC indicator measures the model's capability to distinguish between classification classes. We established that the criteria to identify the performance of a model was the AUC value due to the information it gives about the model.

We divided the data into seven batches to randomly assign a different batch to subsets of the same size. We also tried seven different iterations of the validation process through a performance evaluation for binary variables to calculate the AUC, Accuracy, and Classification error. The suitability of the model was decided in terms of interpretability. The results and information obtained from the models matched the criteria given by related works about what variables can help classify labor turnover.

Case study

Malumex is an outsourcing company located in Ciudad Obregón, Mexico that dedicates about 80 % of its operations to hiring manufacturing operators for food processing companies. The activities of the most workforce hired are manual operations such as packaging and boxing, i.e., operative jobs. The profile of these employees involves a basic educational level and requires no experience or specialized skills. The labor conditions include low salaries, extended work schedules with work shifts of up to 12 hours, and a lack of ergonomics in the facilities. The contract scheme defines an initial work period of 90 days so that the company guarantees that the investment associated with the recruitment process is recovered. The contract is renewed every three or six months, depending on factors such as employee performance.

Figure 2 shows the labor turnover rate of employees that worked with the principal client of Malumex between January 2017 and March 2019. As seen, the turnover percentage tends to increase as high as 50 % despite the implemented strategies to reduce it. The fluctuation of turnover over months makes it difficult to determine patterns. These relatively high turnover rates have had adverse effects on the company, including the high costs of attracting talent, unattainable objectives of the HRM department, and the work overload associated with the lack of the required workforce. This situation has also affected the financial situation of Malumex, as the lack of human talent makes it difficult to close negotiations with new clients. Labor turnover also affects Malumex's relationship with current clients due to the impossibility of maintaining the minimum requirement for personnel operations. The side effects of this situation were the high cost of liquidation of the employees, the penalty imposed by the client, and a reduction of income.

Figure 2 Monthly turnover rate at Malumex in 2017-2019 

The strategies that Malumex has implemented to reduce the labor turnover rate include a bonus for hiring and rewards for the assistance and permanence of employees in the company. Another strategy implemented in the summer of 2017 was to analyze the profile of the successful employee to improve the hiring process. For instance, the minimum hiring age was increased from 18 to 23 years based on the opinion that people under 23 years were unsuitable for this job. Nevertheless, these strategies did not have the desired success. In this context, understanding the profile and behaviors of employees prone to leave the job becomes evident. In the case of Malumex, before the initial contract is completed.

Once we analyzed the situation in Malumex, it was found that the turnover rates present increments that negatively impact the company. Malumex has already identified that many employees leave their job before the 90-day work period. Therefore, analyzing and classifying the employee profiles was crucial for this project. In this sense, the objective of the data mining project was to create a classifier model for employees who resign before 90 days of hiring, as they are not profitable for Malumex.

Data understanding

The initial data set extracted from superNOMINA software contained information on employees from 1998 to March 2019. This software manages information about job characteristics and socio-demographic data. Job applications and digital records of job interviews were also utilized to complement the data set. In general, all these data sources contained missing values.

The final data set included 20 variables and 3735 records (see Table 1). As part of the preparation tasks, new variables and their corresponding values were calculated and included in the dataset: Age and Seniority. These two variables represented the socio-demographic information of employees. A descriptive analysis of SPSS helped to understand the data set better. Table 2 shows measures of the central tendency for these attributes. The average Seniority is 188.50 days, with an s.d. of 422.57, which implies a vast dispersion of data. The median of the distribution is equal to 73 days. We decided to represent the Seniority variable with the median value.

Figure 3 shows the frequency of some categorical and binary variables: Gender, Education, Job experience in Time and Sector, and Job Resignation Reason. We can observe that the male population is higher (58 %) than the female population (42 %). Also, the most frequent education category is Junior High School (65.41 %), while Unjustified absence (33.22 %) is the most common resignation reason. Meanwhile, 48.08 % of the employees have previous experience in factories, and their experience in time is often less than three months (38.62 % of the population).

Figure 3 Frequencies of nominal attributes in the data set utilized 

Data preparation

We first defined two new attributes: Age and Seniority. The values of the second attribute were calculated utilizing information from the available data about periods working at Malumex. The latter attribute was established as the dependent variable since it represents the working time of employees and therefore results helpful to define whether an employee left the job before the initial contract of 90 days (i.e., related to undesirable employee profile).

Variables such as the ID of employees and dates such as date of birth, resignation, and hiring were excluded since these did not contribute to defining employee profiles prone to turnover or were already considered in the values of some other variable (e.g., Age). We also transformed binary and categorical attributes into numeric attributes so that the data mining techniques could be applied. Regarding missing values, the values of Seniority were filled with the median value equal to 73. For the attributes Experience in Time, Experience in the sector, Job resignation reason, Marital status, Progeny, Education, and Location, the missing values were not substituted not to skew the data. Finally, we transformed the dependent variable Seniority into a binary attribute to distinguish the employees who worked less than 90 days from those who worked a more extended period. The categories for this attribute were set as true and false, and the target for the data mining problem was labeled as false.

Modeling

In the first iteration, the techniques listed below were utilized to build classification models using 17 total attributes in the data set. The attribute Job separation reason was excluded since this variable would not be available in a new hiring scenario.

The parameters set for each model are as follows:

A second iteration was carried out using the three models with the highest AUC value (see Table 4) to evaluate whether the attribute with a high correlation with the dependent variable "Job separation reason" had a significant impact, where an automatic feature selection and generation was established for finding the best feature sets. The parameters for the second iteration are given in Table 3.

Evaluation

In the first iteration (Table 4), the results showed that the models tested had an accuracy within the range of .613 to .621, where SVM was the model that presented the highest accuracy. Nevertheless, the RF presented an AUC value equal to .625 while the rest of the models had an AUC around .500 - .550; an AUC near .500 means the model could not distinguish between classes. With those evaluations and interpretations, the model with the best performance on the first iteration was the RF.

We utilized the three models with the highest AUC value for the second iteration. We tested an automatic feature selection of the training and evaluation subsets to evaluate the impact on the performance of the AUC value. The results (see Table 5) showed an accuracy value higher than the first iteration for the three models, but only the SVM had a performance above .500 on the AUC value. The accuracy and classification error values did not present significant differences (as the variation between iterations was less than 0.100, and it did not get close to 1).

Although these two models (RF with no automatic feature selection and SVM with automatic feature selection) are the ones with an AUC above .500, it can be considered either RF or SVM as the best models. The weights of the variables selected for both models are presented in Table 6.

The SVM algorithm performed better when an automatic feature selection and generation were set, with an accuracy of 70.1 % and AUC of 0.645. This model is not wholly appropriate for profiling new employees, but it can be beneficial to track and predict the performance of retired personnel. The reason is that most attributes refer to job characteristics rather than socio-demographic information. The RF algorithm presented an accuracy of 62 % and AUC of 0.625. This model could help classify the future performance of new prospects and candidates as it includes attributes that refer to the socio-demographic profile of workers.