“How to Make Them Stay?”: Diverse Counterfactual Explanations of

Employee Attrition

Andr

e Artelt

1,2 a

and Andreas Gregoriades

3 b

Faculty of Technology, Bielefeld University, Bielefeld, Germany

Department of Electrical and Computer Engineering, University of Cyprus, Nicosia, Cyprus

Department of Management, Entrepreneurship and Digital Business, Cyprus University of Technology, Limassol, Cyprus

Keywords:

Employee Attrition Prediction, Counterfactual Explanations, Explainable Machine Learning.

Abstract:

Employee attrition is an important and complex problem that can directly affect an organisation’s competitive-

ness and performance. Explaining the reasons why employees leave an organisation is a key human resource

management challenge due to the high costs and time required to attract and keep talented employees. Busi-

nesses therefore aim to increase employee retention rates to minimise their costs and maximise their perfor-

mance. Machine learning (ML) has been applied in various aspects of human resource management including

attrition prediction to provide businesses with insights on proactive measures on how to prevent talented em-

ployees from quitting. Among these ML methods, the best performance has been reported by ensemble or

deep neural networks, which by nature constitute black box techniques and thus cannot be easily interpreted.

To enable the understanding of these models’ reasoning several explainability frameworks have been proposed

to either explain individual cases using local interpretation approaches or provide global explanations describ-

ing the overall logic of the predictive model. Counterfactual explanation methods have attracted considerable

attention in recent years since they can be used to explain and recommend actions to be performed to obtain

the desired outcome. However current counterfactual explanations methods focus on optimising the changes

to be made on individual cases to achieve the desired outcome. In the attrition problem it is important to be

able to foresee what would be the effect of an organisation’s action to a group of employees where the goal is

to prevent them from leaving the company. Therefore, in this paper we propose the use of counterfactual ex-

planations focusing on multiple attrition cases from historical data, to identify the optimum interventions that

an organisation needs to make to its practices/policies to prevent or minimise attrition probability for these

cases. The proposed technique is applied on an employee attrition dataset, used to train binary classiﬁers.

Counterfactual explanations are generated based on multiple attrition cases, thus, providing recommendations

to the human resource department on how to prevent attrition.

1 INTRODUCTION

Employees constitute one of the most valuable assets

in any organization. Therefore, the optimum way to

manage this resource signiﬁcantly improves organi-

sational performance and competitiveness while also

assists in obtaining organisation’s objectives. Hu-

man Resource (HR) management departments there-

fore engage in activities that aim to unleash employ-

ees’ full potential and maximise their productivity.

Typical HR activities include the process of selec-

tion and recruitment, performance management, em-

ployee well being and satisfaction, training and de-

https://orcid.org/0000-0002-2426-3126

https://orcid.org/0000-0002-7422-1514

velopment. Due to the excessive cost associated with

the selection, recruitment and training of employees

before they become productive, HR management is

constantly striving to keep their employees satisﬁed

since the retention of talented employees is crucial to

any company’s success. Retention involves the sys-

tematic effort of creating a working environment that

satisﬁes employees’ needs by implementing appropri-

ate policies and practices. Identifying optimum poli-

cies is a key problem in HR management and is the fo-

cus of this work. To address this issue, it is essential

to understand what causes employees’ attrition that

refers to the situation when an employee leaves the

business either voluntarily or involuntarily. In the for-

mer case an employee makes a personal decision to

532

Artelt, A. and Gregoriades, A.

“How to Make Them Stay?”: Diverse Counterfactual Explanations of Employee Attrition.

DOI: 10.5220/0011961300003467

In Proceedings of the 25th International Conference on Enterprise Information Systems (ICEIS 2023) - Volume 1, pages 532-538

ISBN: 978-989-758-648-4; ISSN: 2184-4992

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

leave the company and in the latter an employee is

forced to resign due to low performance or not de-

sired skills. Voluntary turnover (attrition) tends to

relate with more skilled and talented employees and

thus the loss for an organisation is signiﬁcant in such

cases due to loss of expertise which in certain cases

might create operational issues. Therefore, voluntary

turnover can have direct and indirect effects to an or-

ganisation through increased hiring and training costs,

reduced productivity, proﬁts and employee morale.

Therefore, employees’ intention to leave an organi-

sation is a widely studied topic, with many studies

investigating different factors that positively or neg-

atively inﬂuence it. Such studies utilise question-

naires completed either by active employees to mea-

sure their turnover intention, or data from employees

that provided their resignation notice explaining their

motivations behind the decision to leave. Utilisation

of turnover intention data however is considered the

most feasible approach in attrition prediction since

detailed resignation data is often unavailable due to

privacy policies. Such studies test different hypothe-

ses or provide new knowledge regarding the way dif-

ferent personal and organisational factors affect reten-

tion/attrition, mainly through statistical signiﬁcance

tests via regression and analysis of variance. How-

ever, such studies provide limited insights or recom-

mendations on what organisations need to do to pre-

vent attrition under different circumstances since their

insights focus on how each factor affects attrition and

not on what state these factors should be in collec-

tively to prevent attrition. Such optimum states of

attrition factors can be identiﬁed through interpreta-

tions of machine learning models trained on historical

data and interrogated through explanation methods

that provide a recommendation what to do in order

to increase retention. Machine learning has been ap-

plied extensively in different prediction tasks includ-

ing attrition (Zhao et al., 2018) due to its improved

prediction performance compared to traditional tech-

niques. However, black box techniques tend to out-

perform the accuracy of interpretable methods such as

decision trees. Thus, different explainability methods

have been proposed to shade light into the logic of en-

semble or deep neural network models, by identifying

patterns in their reasoning (Molnar, 2020; Guidotti

et al., 2018; Linardatos et al., 2020). Counterfactual

explanations (Wachter et al., 2017) (CE) are a popu-

lar technique for interpreting black box models that

enable the identiﬁcation of the optimum changes to

be made to a model’s feature values to obtain the de-

sired outcome. The approach presented in this paper

aims to identify the minimum consistent feature-value

changes that reverse attrition prediction for a group of

employees rather than individual cases and thus pro-

vide valuable recommendations to management on

what to focus on maximising retention. The identiﬁ-

cation of recommendations based on a group of cases

(employee attrition) is what differentiates the method

proposed in this paper from other studies that apply

counterfactual explanations to the employee attrition

problem. The data that we utilise to demonstrate the

proposed approach is the IBM dataset (IBM, 2020)

that is publicly available.

2 LITERATURE REVIEW

Employee attrition prediction using ML has gained

signiﬁcant attention in recent years, with scholars

utilising a wide range of ML techniques on work-

related factors to predict employee turnover inten-

tion (Zhao et al., 2018). Indicatively the follow-

ing ML techniques have been applied to the em-

ployee attrition prediction problem: Support vector

machines, Decision trees, Random Forest, XGBoost,

Logistic Regression, N

aıve Bayes, Adaptive Boost-

ing, K-nearest neighbors, Artiﬁcial neural network,

and Light Gradient Boosting. To the best of our

knowledge the best performance is reported with the

SVM and LGB. Prevalent features used to train such

models relate to individual and organisational factors

such as number of promotions, salary, last evaluation,

time spent in company, working conditions, work-

ing hours, employee-related factors such as age, gen-

der, job satisfaction, work life balance, emotional ex-

haustion, growth potential, and marital status. These

factors have been identiﬁed in previous works (Kang

et al., 2021; Le et al., 2022) that examined how indi-

vidual and organisational factors affect employee at-

trition. The effect of the combination of these factors

varies. For instance, factors such as time in service

showed that employees who have worked for longer

at a given organization are less likely to leave. While

longer time before employees obtain organizational

tenure positively mitigates to employees’ intentions

to leave. Sex and education have also been identi-

ﬁed as inﬂuencing factors however there is no con-

sensus on how each factor affects attrition. Organiza-

tional factors on the other hand are factors controlled

by the organisation and thus could be altered to im-

prove retention. Such factors are job satisfaction that

is strongly and signiﬁcantly related to employees’ in-

tentions to leave, with employees who feel higher per-

sonal accomplishment in their job being less likely

to leave their organization. Overloading on the other

hand results in emotional, physical, and mental ex-

haustion and thus increases employees’ intentions to

“How to Make Them Stay?”: Diverse Counterfactual Explanations of Employee Attrition

533

leave. Employees who think of their work as mean-

ingful are less likely to leave. Finally, employees who

understand the relationship of their job to an organiza-

tions’ goals are also less likely to leave their organiza-

tion. Since they consider themselves as part of a team

with a common objective. Several organisational poli-

cies have been reported to improve retention such as

salary, family-friendly policies, training, skill devel-

opment, and diversity management. Speciﬁcally, re-

wards such as salary, increase employees’ motivation

and satisfaction which makes them less likely to leave

their organizations in contrast to employees that are

not satisﬁed with their salary. In addition, employees

who are not rewarded for their good performance are

more likely to leave the organization. On the other

hand, balancing work and family responsibilities re-

duce employees’ turnover. While training and devel-

opment opportunities are found to mitigate turnover

but can also increase turnover since employees be-

come more attractive by competitors. Working envi-

ronment on the other hand, such as good relationships

with co-workers and fairness in organisational proce-

dures reduce intention to leave, while organizations

that support creativity and innovation are signiﬁcantly

related to lower turnover intentions due to the asso-

ciated sense of meaning and pleasure from the work

done. While, employees who feel supported by their

organization are less likely to leave the organization

since they consider that the organisation takes care of

them. Finally, a meta-analytic review of voluntary at-

trition (Haldorai et al., 2019) found that the strongest

predictors of employee turnover are, age, pay, and

job satisfaction. Other studies highlight that “anti-

social” working hours, work life conﬂict, emotional

exhaustion, work overload, working environment, ca-

reer progression and community ﬁt strongly inﬂuenc-

ing turnover intention.

3 COUNTERFACTUAL

EXPLANATIONS

In predictive modelling it is important to have good

accuracy of what is predicted but also high inter-

pretability. The former is essential to be able to make

actionable decisions while the latter increases the con-

ﬁdence & trust in the predictions and understanding

of the model’s logic. However, most models that

produce high predictive accuracy are also less inter-

pretable and vice versa. Several (explanation) meth-

ods have been developed to increase the interpretabil-

ity of black box models (Molnar, 2020), however, no

practical technology has yet emerged for explainable

AI (Guidotti et al., 2018; Linardatos et al., 2020).

This is due to the complexity of the matter with ex-

planations having to be both statistically sound and

comprehensible by stakeholders. Many approaches to

the explainability problem focus on the global logic of

a black box model through an associated interpretable

classiﬁer such as decision tree that mimics the behav-

ior of the black box model. These methods are model

dependent. A different branch of work on model-

independent (agnostic) techniques for understanding

black box models’ behaviour is by using the classi-

ﬁers’ predictions to generate explanations. These are

referred as post hoc since the technique is applied af-

ter model training. Such model agnostic techniques

are further categorised into local and global expla-

nation methods. The local focusing on speciﬁc in-

stances and global on the whole set of instances. A

few recent methods that are model-agnostic, such as

LIME (Ribeiro et al., 2016) obtain a local explanation

for a decision outcome by learning an interpretable

model from querying randomly perturbed versions of

a given instance on a black box model. Such post-

hoc methods for extracting local explanations from

each model prediction have been attracting much at-

tention. Counterfactual explanation (Wachter et al.,

2017; Verma et al., 2020) is a post-hoc local expla-

nation method that show why the undesirable predic-

tions emerged and what needs to be changed in the in-

put to obtain the desired results. Thus a counterfactual

explanation describes a causal situation in the form: If

X had occurred, Y would have occurred. Counterfac-

tual explanations can be used to explain predictions

of individual instances and determine required input

to obtain opposite results.

To generate counterfactuals researchers use

optimization-based approaches. Therefore, once you

have a trained classiﬁer the goal of the optimiser is

to ﬁnd a counterfactual with the shortest distance to

a case with the desired output (non attrition). Such

counterfactual can be obtained by solving an opti-

mization problem.

Various local explanation methods however have

been criticized for not being robust (Artelt et al.,

2021; Hancox-Li, 2020; Mishra et al., 2021) or that

they might fail to explain the global behavior of com-

plex models (Slack et al., 2021). Thus, such meth-

ods have limited applications in problems such as em-

ployee attrition that require the identiﬁcation of the

optimum changes to be made by the HR department

to prevent attrition. For such a problem, multiple in-

stances of employees that quit their jobs need to be

considered. To address the above issue, several meth-

ods for tackling the multiple local explanations prob-

lem have been proposed. For instance, AReS (Rawal

and Lakkaraju, 2020),which is a global summary of

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

534

actions expressed in rules and MAME (Rama-murthy

et al., 2020) and GIME (Gao et al., 2021) that are

global summaries of general local explanation meth-

ods (e.g., LIME (Ribeiro et al., 2016)). Recently,

another approach (Kanamori et al., 2022) for simul-

taneously computing counterfactual explanations for

different groups of individuals was proposed that is

similar to the approach presented in this paper.

4 METHODOLOGY

We propose a counterfactual explanation methodol-

ogy for computing a recommendation (i.e. explana-

tion) of what the HR department should do to in-

crease retention of employees. Essentially the goal

is to ﬁnd the changes the HR departments must do

to their policy such that many (if not all) of the em-

ployees that initially intended to leave the company

change their minds. For this purpose we assume that

we have a set of employees where an attrition clas-

siﬁer h : R

→ {−1, 1} predicts that they quit their

jobs. We denote this set of employees as D and as-

sume that each employee is represented by some real-

valued feature vector⃗x

∈ R

4.1 A Single Explanation

We are looking for the changes the HR department

must make

⃗

∈ R

which, if applied to the employ-

ees ⃗x

, change the prediction of the attrition classiﬁer

h(·) from attrition to retention. We call such a

⃗

a counterfactual explanation of attrition. We phrase

this as the following optimization problem – since we

are looking for a “simple” (i.e. cost sensitive) rec-

ommendation of changes

⃗

, we aim to minimize the

number and magnitude of changes in

⃗

by using the

1-norm

min

⃗

∈R

∥

⃗

∥

+C ·

∑

⃗x

∈D

ℓ



h(⃗x

⃗

)



(1)

where ℓ(·) denotes a suitable loss function penaliz-

ing attrition predictions (i.e. penalty if the change

⃗

does not ﬂip the output of the model for an employee).

Suitable loss functions might be mean-squared er-

ror or cross-entropy loss. The regularization strength

C > 0 allows us to balance between the two objectives

of having a “simple” recommendation of actions and

a recommendation that works for as many employees

as possible.

The 1-norm treats all features as equally important and

costly – this can be changed by using a weighted 1-norm

instead.

While our formalization Eq. (1) is completely

model-agnostic – i.e. it can be applied to any attri-

tion classiﬁer h(·), it might not be the most efﬁcient

formalization for every situation. In particular if we

have knowledge about h(·) that could be exploited for

a more efﬁcient (e.g. faster) computation of the rec-

ommendation

⃗

, this can used. In the following,

we investigate the special case of linear classiﬁers:

In case of a linear classiﬁer h(⃗x) = sign



⃗w

⊤

⃗x + b



such as linear-SVM, logistic regression, etc., we can

rewrite Eq. (1) as the following linear program (LP):

min

⃗

∈R

∥

⃗

∥

+C ·

∑

s.t. y



⃗w

⊤

⃗

+⃗w

⊤

⃗x

+ b



≥ 0 − ξ ∀⃗x

∈ D

ξ ≥ 0 ∀i

(2)

where C > 0 is again a hyperparemeter that allows us

to balance between the two objectives. In practice it

might be necessary to try different values for C in or-

der to ﬁnd a practically satisfying/useful recommen-

dation

⃗

. Note that linear programs (LP) are special

instances of convex optimization problems that can be

solved very efﬁciently (Boyd et al., 2004).

4.2 A Set of Diverse Explanations

As already noted by (Kanamori et al., 2022), there

often might not exist a single change

⃗

that is ap-

plicable to all employees in D. Furthermore, there

might also exist several different changes

⃗

that

work equally well (and maybe also work for differ-

ent subgroups of employees), we therefore propose an

extension to our formalization from Section 4.1 that

computes not a single recommendation of changes

⃗

but a set of different & diverse changes

⃗

so that the

decision makers are provided with a list of possible

actions on how to increase retention rate. Therefore,

the HR manager can choose the action that is more

suitable to his/her case.

We propose an iterative method that computes a

set of highly diverse changes

⃗

. We deﬁned diver-

sity by means of the number of overlapping features.

For instance, highly diverse explanations should use

& change completely different features. For this pur-

pose, we need a mechanism for excluding already

used features from future changes

⃗

. We can mod-

ify our previous optimization problems not to use any

black-listed features F – i.e. features already used

by previous recommendations of changes – by intro-

ducing a diagonal matrix M ∈ R

d×d

and replacing all

occurrences of

⃗

(3)

“How to Make Them Stay?”: Diverse Counterfactual Explanations of Employee Attrition

535

The diagonal matrix M is then deﬁned as follows:

(M)

i,i

(

0 if i ∈ F

1 otherwise

(4)

The effect of Eq. (3) is that it sets all changes in black-

listed features F to zero – i.e. it basically removes

forbidden changes and the solver is required to ﬁnd

other changes

⃗

that yield a feasible solution. Note

that Eq. (3) does not change the computational com-

plexity of the original optimization problems Eq. (1)

and Eq. (2). However, it can happen that for a set of

black-listed features F no feasible solution exists –

i.e. no allowed changes will change the attrition pre-

diction.

We introduce the following notation: CF

(D,F )

computes the solution to one of the previous optimiza-

tion problems (i.e. depending on h(·) either Eq. (1)

or Eq. (2)) under the additional constraint that no

black-listed features F must be used in the ﬁnal solu-

tion. The complete procedure is described in pseudo-

code in Algorithm 1.

Algorithm 1: Computation of Diverse Counterfactual Ex-

planations of Attrition.

Input: Set of employees D where attrition is pre-

dicted; k ≥ 1: number of diverse counterfactual ex-

planations of attrition; attrition classiﬁer h(·)

Output: Set of diverse counterfactuals R = {⃗x

}

1: F = {} ▷ Initialize set of black-listed features

2: R = {} ▷ Initialize set of diverse counterfactuals

3: for i = 1, .. ., k do ▷ Compute k diverse

counterfactuals

⃗

= CF

(D,F ) ▷ Compute next

recommendation of changes

5: R = R ∪ {

⃗

}

6: F = F ∪ { j | (

⃗

)

̸= 0} ▷ Update set of

black-listed features

7: end for

4.3 Related Work

The authors of (Kanamori et al., 2022) deal with a

very similar problem, which they call “group-wise

counterfactual explanations”. However, the authors

do not try to ﬁnd a single explanation applicable to

as many as possible employees like we do in Sec-

tion 4.1, but instead propose an algorithm/heuristic

called “counterfactual explanation tree” for partition-

ing the employees into groups for which a single ap-

plicable explanation (i.e. recommendation of changes

⃗

) is computed. However, their proposed method

is computationally expensive and the algorithm for

building the counterfactual explanation tree is just a

heuristic that comes without any formal guarantees.

Furthermore, their algorithm can not be easily ex-

tended or customized – e.g. adding constraints to the

explanations or computing multiple explanations –

whereas our optimization as illustrated in Section 4.1

can be extended and reused in many different ways as

we do in Section 4.2 for computing a set of diverse

changes

⃗

5 CASE STUDY

To illustrate the application of the proposed multi in-

stance counterfactual method in the employee attri-

tion problem the following data-set is utilised.

5.1 Dataset

The IBM human resource dataset (IBM, 2020) con-

tains 35 features for 1467 unique employees. The

dataset contains human resources properties such as

age, education, gender, promotion, education and

rate. The question we want to answer using this data

is “what does the HR department need to do to pre-

vent employees from leaving the organisation”. To

address this question all the attrition cases need to

be considered and the optimum solution that will sat-

isfy all these cases is identifying using the proposed

method. The method utilise only features that can be

changed such as salary, years before promotion, busi-

ness travel, environmental conditions, etc. Addition-

ally, for this example we utilise only numerical fea-

tures since they are easier to manipulate and thus pro-

vide proof of concept. However categorical features

can also be utilised after undergoing one-hot encoding

or when converted to ordinal variables (e.g. education

can be expressed as an ordinal variable with high val-

ues indicating education of PhD level). Features that

can not be changed such as age, gender etc are not

utilised. However these features can be utilised when

hiring employees, so as to target employees that are

more likely to stay at the organisation. The method

can also be used with features that are generated from

combinations of existing features, for example a new

feature “job hopper” that refers to people that change

jobs regularly to increase their salary and thus may

have a higher probability of leaving the company can

also be generated and utilised in the analysis. This can

be estimated from the number of companies the em-

ployees worked before divided by the total working

years.

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

536

5.2 Setup

We implement logistic regression, random forest

and XGBoost as attrition classiﬁers using the IBM

dataset. To account for the differences in job roles

and departments the analysis can be performed by

focusing on cases refering to different departments

such as ”Research and development” or ”Sales”. This

could give more speciﬁc recommendations that can

relate better to the department in question.For this

case study we focused in the Research and Develop-

ment department and the features that are utilized are:

EnvironmentSatisfaction, JobInvolvement, JobSatis-

faction, MonthlyIncome, PercentSalaryHike, YearsIn-

CurrentRole, YearsSinceLastPromotion, YearsWith-

CurrManager. These features have been known to

affect attrition based on the literature, thus we aimed

to identify how the HR department can alter the inten-

tion of employees by changing individual or combi-

nations of these factors. The contribution is the inten-

sity of the change and the combination of the features

to achieve the desired outcome. All classiﬁers have

been optimised using hyper-parameter tuning. Data

is split into train and test set for ﬁnding the optimal

hyper parameters. Furthermore, data is standardized

and random under-sampling of the majority class is

performed in order to avoid any biases in the classi-

ﬁer due to class imbalance.

5.3 Results

We observe that our method (Section 4) computes

several reasonable recommendations (of different

complexity) on how to reduce employee attrition

and provided valuable insights on employee turnover

which is consistent with the literature. Speciﬁcally,

the following recommendations refer to turnover in-

stances and indicate the features that must be manip-

ulated and by how much to prevent employees from

leaving the company:

1. If employees would have had an increase in Per-

cent Salary Hike of approx. 40%, attrition would

be unlikely.

2. If employees would have had approx. 5 years less

since their last promotion, attrition would be un-

likely.

3. If employees would have had an increase in Per-

centage Salary Hike of approx. 20%, AND had in

an increase in job satisfaction by approx. 50% ,

attrition would be unlikely.

These recommendations highlight the importance

of salary, job satisfaction and promotion as key fac-

tors for preventing employee turnover, which abide

with the theory or retention. Speciﬁcally, the results

from the example case study show that to achieve

this goal the company needs to increase the salary

of selected employees or change its promotion pol-

icy so as to motivate employees by promoting them

earlier. However, given that the recommended salary

increase is large, companies could apply the recom-

mendation on talented employees that can not af-

ford to loose. Similar actions are recommended by

(Kanamori et al., 2022), thus highlighting the impor-

tance of salary (monthly income) as key factors for re-

tention, while also provides conﬁdence in our prelim-

inary results. In addition, as highlighted in the litera-

ture the method also identiﬁes that job satisfaction is a

key property for preventing attrition if combined with

salary increase. Increase in job satisfaction can be re-

alised through improved working conditions such as

renovation of the working environment, bonding ac-

tivities, organised events, complimentary food etc.

In contrast to other emplainable ML methods such

as SHAP (Lundberg and Lee, 2017) that provide

global explanations and specify the factors that in-

ﬂuence the class variable (attrition), the proposed

method can go one step further by providing speciﬁc

recommendations of business policy changes to re-

vert attrition of speciﬁc employees. These provide

actionable business recommendations that in contrast

to SHAP provide speciﬁc policy changes that are op-

timised on multiple instances (for instance employees

in speciﬁc department that is key to the competitive-

ness of the company and their skills cannot easily be

replenished) and explain what the management needs

to do to prevent attrition of these employees.

6 CONCLUSIONS

The method proposed in this work utilise counterfac-

tual explanations but in contrast to other similar stud-

ies utilise multiple instances to generate recommen-

dations that are actionable by businesses. This is im-

portant in problems such as employee attrition where

speciﬁc HR actions on what needs to be done to pre-

vent talented employee attrition should be based on

multiple instances rather than individual employees

that left the company. In the case of a single em-

ployee attrition prevention, the company needs only

to consider the properties of an individual to derive

actions on how to prevent him/her from quitting, how-

ever, in reality the company needs a general policy to

reduce attrition for all turnover cases due to the neg-

ative effect that attrition has on business performance

and competitiveness. The method is applied on the

IBM attrition dataset showing policy changes that are

“How to Make Them Stay?”: Diverse Counterfactual Explanations of Employee Attrition

537

valid since they abide with the literature. Future work

will focus on providing weights to the features that

are manipulated so as for the method to provide rec-

ommendations ranked based on how easy these can be

implemented. For instance salary increase by 30 per-

cent might be more difﬁcult to realise than increase in

company’s environmental conditions. In addition the

method will be extended to eliminate automatically

infeasible solutions by enabling the user to provide

range of feasible feature values. Future work will also

focus on comparing the results of our method with

similar work under different scenarios.

ACKNOWLEDGEMENTS

We gratefully acknowledge funding from the VW-

Foundation for the project IMPACT funded in the

frame of the funding line AI and its Implications for

Future Society.

REFERENCES

Artelt, A., Vaquet, V., Velioglu, R., Hinder, F., Brinkrolf,

J., Schilling, M., and Hammer, B. (2021). Evalu-

ating robustness of counterfactual explanations. In

2021 IEEE Symposium Series on Computational In-

telligence (SSCI), pages 01–09. IEEE.

Boyd, S., Boyd, S. P., and Vandenberghe, L. (2004). Convex

optimization. Cambridge university press.

Gao, J., Wang, X., Wang, Y., Yan, Y., and Xie, X. (2021).

Learning groupwise explanations for black-box mod-

els. In IJCAI, pages 2396–2402.

Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Gian-

notti, F., and Pedreschi, D. (2018). A survey of meth-

ods for explaining black box models. ACM Comput.

Surv., 51(5).

Haldorai, K., Kim, W. G., Pillai, S. G., Park, T. E., and

Balasubramanian, K. (2019). Factors affecting ho-

tel employees’ attrition and turnover: Application of

pull-push-mooring framework. International Journal

of Hospitality Management, 83:46–55.

Hancox-Li, L. (2020). Robustness in machine learning

explanations: does it matter? In Proceedings of

the 2020 conference on fairness, accountability, and

transparency, pages 640–647.

IBM (2020). Ibm hr analytics employee.

https://www.kaggle.com/pavansubhasht/ibm-hr-

analytics-attrition-dataset.

Kanamori, K., Takagi, T., Kobayashi, K., and Ike, Y.

(2022). Counterfactual explanation trees: Transparent

and consistent actionable recourse with decision trees.

In International Conference on Artiﬁcial Intelligence

and Statistics, pages 1846–1870. PMLR.

Kang, I. G., Croft, B., and Bichelmeyer, B. A. (2021). Pre-

dictors of turnover intention in us federal government

workforce: Machine learning evidence that perceived

comprehensive hr practices predict turnover intention.

Public Personnel Management, 50(4):538–558.

Le, H., Lee, J., Nielsen, I., and Nguyen, T. L. A. (2022).

Turnover intentions: the roles of job satisfaction and

family support. Personnel Review, (ahead-of-print).

Linardatos, P., Papastefanopoulos, V., and Kotsiantis, S.

(2020). Explainable ai: A review of machine learn-

ing interpretability methods. Entropy, 23(1):18.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. Advances in neural

information processing systems, 30.

Mishra, S., Dutta, S., Long, J., and Magazzeni, D.

(2021). A survey on the robustness of feature impor-

tance and counterfactual explanations. arXiv preprint

arXiv:2111.00358.

Molnar, C. (2020). Interpretable machine learning. Lulu.

com.

Rawal, K. and Lakkaraju, H. (2020). Beyond individualized

recourse: Interpretable and interactive summaries of

actionable recourses. Advances in Neural Information

Processing Systems, 33:12187–12198.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ” why

should i trust you?” explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

international conference on knowledge discovery and

data mining, pages 1135–1144.

Slack, D., Hilgard, A., Lakkaraju, H., and Singh, S. (2021).

Counterfactual explanations can be manipulated. Ad-

vances in Neural Information Processing Systems,

34:62–75.

Verma, S., Dickerson, J., and Hines, K. (2020). Counter-

factual explanations for machine learning: A review.

arXiv preprint arXiv:2010.10596.

Wachter, S., Mittelstadt, B., and Russell, C. (2017). Coun-

terfactual explanations without opening the black box:

Automated decisions and the gdpr. Harv. JL & Tech.,

31:841.

Zhao, Y., Hryniewicki, M. K., Cheng, F., Fu, B., and Zhu,

X. (2018). Employee turnover prediction with ma-

chine learning: A reliable approach. In Proceedings

of SAI intelligent systems conference, pages 737–758.

Springer.

ICEIS 2023 - 25th International Conference on Enterprise Information Systems

538