TABLE OF CONTENTS

TABLE OF CONTENTS............................................................................................ 1
LIST OF ABBREVIATIONS..................................................................................... 3
INTRODUCTION....................................................................................................... 4
CHAPTER 1: DATA OVERVIEW & DESCRIPTIVE STATISTICS................... 6
1.1 Overview of the Dataset.....................................................................................6
1.2 Descriptive Statistics..........................................................................................6
CHAPTER 2: EXPLORATORY DATA ANALYSIS (EDA)................................... 8
2.1 Univariate Analysis............................................................................................8
2.1.1 Categorical Variables.................................................................................8
2.1.2. Numerical Variables................................................................................. 9
2.2 Bivariate & Multivariate Analysis................................................................... 11
2.2.1 Overview of Employee Performance Rating........................................... 11
2.2.2 Employee Performance by Department & Job Involvement...................13
2.2.3 Employee Performance & Workplace Factors: Job Level, Work-Life
Balance & Job Satisfaction...............................................................................15
2.2.4 Employee Performance & Career Growth, Workload: Promotions, Salary
Hikes & Overtime.............................................................................................17
2.3 Correlation Analysis.........................................................................................19
2.4 Check Skewness & Kurtosis............................................................................ 20
CHAPTER 3 : DATA PRE-PROCESSING............................................................ 21
3.1 Check Missing Value........................................................................................21
3.2 Handle Outliers................................................................................................ 22
3.3 Categorical Data Conversion........................................................................... 23
3.4 Feature Transformation.................................................................................... 24
3.5 Feature Selection..............................................................................................25
CHAPTER 4: MACHINE LEARNING MODEL CREATION & EVALUATION
26
4.1. Data Preparation..............................................................................................26
4.2. Model Selection.............................................................................................. 27
4.2.1. Support Vector Machine (SVM).............................................................27
4.2.2. Random Forest (RF)............................................................................... 28
4.2.3. Balanced Random Forest (BRF).............................................................29
4.2.4. Artificial Neural Network - ANN (MLP Classifier).............................. 30
4.3. Model Implementation.................................................................................... 30
4.3.1. Handling Class Imbalance...................................................................... 30
4.3.2. Model Implementation........................................................................... 32
4.3.2.1. Support Vector Machine (SVM).................................................... 32
4.3.2.2. Random Forest (RF).......................................................................33

1
 4.3.2.3. Balanced Random Forest (BRF).................................................... 34
4.3.2.4. Artificial Neural Network - ANN (MLP Classifier)...................... 35
4.4. Model Evaluation............................................................................................ 36
4.4.1. Support Vector Machine (SVM).............................................................36
4.4.2. Random Forest (RF)............................................................................... 37
4.4.3. Balanced Random Forest (BRF).............................................................38
4.4.4. Artificial Neural Network - ANN (MLP Classifier).............................. 38
4.5. Model Comparison..........................................................................................40
CHAPTER 5: KEYS FINDINGS & RECOMMENDATIONS............................. 42
5.1. Key Findings................................................................................................... 42
5.1.1. Employee Environment Satisfaction...................................................... 42
5.1.2. Last Salary Hike Percentage...................................................................42
5.1.3. Years Since Last Promotion....................................................................42
5.2. Strategies to Improve Employee Performance................................................43
CONCLUSION.......................................................................................................... 44
REFERENCE.............................................................................................................45
APPENDICES............................................................................................................50
 LIST OF ABBREVIATIONS

No. Abbreviation Explanation

ANN Artificial Neural Network

BRF Balanced Random Forest

IQR Interquartile Range

MLP Multi-layer Perceptron

Q-Q Quartile - Quartile

SMOTE Synthetic Minority Over-sampling Technique

SVC Support Vector Classifier

SVM Support Vector Machine

WLB Work-life balance

 INTRODUCTION
According to Muslih (2022), performance assessment is a powerful tool
assisting managers in making promotion decisions, employee training, and violation
sanctions. Empirical evidence suggests that organizations prioritizing performance
management tend to outperform their peers financially (McKinsey). However, to
ensure the transparency and accuracy of the assessment, there is a need for a critical
approach and effective support system that is backed up by data and suitable
collection methods.
Aiming at building a practical performance assessment system, data analysis
would be an irreplaceable component. In general, data analysis would not only
provide valuable insights for business performance improvement in the short term but
also suggest predictive models, and help mitigate human resource risks in the long
term. First, based on data, companies can monitor performance actively, identify
trends and patterns, and establish clear performance metrics. This approach enables
personalized training and development programs, and boosts employee satisfaction,
leading to increased productivity and efficiency. Moreover, predictive analytics helps
in forecasting future performance and managing talent effectively, ensuring that
organizations can proactively address potential issues and capitalize on opportunities.
Therefore, this research will aim to dissect employee data to identify key drivers of
performance, construct a predictive model for performance ratings, and provide
actionable recommendations for improvement.
This research is guided by three primary questions. Firstly, what are the key
drivers of employee performance in contemporary organizational settings? This
question seeks to identify the most influential factors affecting employee
performance. Secondly, how can machine learning models be effectively used to
predict employee performance based on historical data? This involves developing a
predictive model that leverages data analytics to forecast performance trends. Lastly,
what recommendations can be derived from data analysis to improve employee
performance and enhance organizational productivity? This question aims to provide
actionable insights and strategies for enhancing employee performance, grounded in
the findings from the predictive model.
 The goal of this research will be addressed by delivering three main objectives
deliberately. Firstly, it seeks to identify the key drivers of employee performance,
focusing on both quantitative and qualitative factors that significantly influence
individual and organizational outcomes. Secondly, the study aims to develop a
predictive model for employee performance using advanced machine learning
techniques, which will enable organizations to forecast performance trends and make
informed decisions about talent development and resource allocation. The last
objective is to provide actionable recommendations for improving employee
performance based on the insights derived from the predictive model, thereby
supporting organizations in designing targeted interventions and strategies to enhance
workforce productivity and engagement.
This research contributes to the existing body of knowledge on employee
performance by identifying key drivers, developing a predictive model, and providing
actionable recommendations. It builds upon previous studies that highlight the
importance of factors such as workplace environment, employee satisfaction,
management standards, and training in influencing employee performance (Frontiers,
2022; Dergipark, 2022). The predictive model developed in this study enhances the
ability of organizations to forecast performance trends, allowing for proactive talent
management and strategic decision-making. Furthermore, by emphasizing
data-driven insights, this research supports the creation of personalized development
programs and targeted interventions, which can lead to improved job satisfaction and
retention (Alwardi, 2023). Overall, this study contributes to the development of more
effective employee performance management strategies, aligning with broader
organizational goals and fostering a high-performing organizational culture (Mesiya,
2022).
 CHAPTER 1: DATA OVERVIEW & DESCRIPTIVE STATISTICS
1.1 Overview of the Dataset
This paper makes use of an employee performance dataset consisting of 1,200
observations and 28 features. The dataset encompasses both numerical and
categorical variables capturing various aspects of employee performance, workplace
attributes, and individual characteristics. The primary objective of this analysis is to
leverage these features to predict employee performance ratings, which serve as the
target variable.
Among 19 quantitative variables, 11 are numerical (e.g., age, hourly rate,
years of experience) and 8 are ordinal (e.g., job satisfaction, work-life balance).
Additionally, the dataset contains 8 qualitative features, which represent categorical
attributes, such as department, marital status, job role, and among others. The
"EmpNumber", an alphanumeric identifier, will be excluded as it does not contribute
to employee performance prediction.
1.2 Descriptive Statistics
Figure 1.1. Summary statistics of the dataset

Source: The authors (2025)

 The average age of employees is 37 (36.92) years old (SD = 9.09), ranging
from 18 to 60, which shows that the company employs both early-career and
experienced workers. Commute distances vary significantly (Avg = 9.17 km, SD =
8.18), with commutes ranging from 1 to 29 km. This indicates that while some
employees live closer to the workplace, other employees have much longer
commutes.
This dataset contains metrics related to engagement and job satisfaction as well.
Overall, the employee involvement score lies at 2.73, which is moderate given the
range of from 1 to 4. The average level of job satisfaction is noted at 2.73 as well,
similar to that of the relationship satisfaction among employees. Due to low standard
deviations (around 1.1), it is clear that most employees are within a similar level of
fulfillment when it comes to these parameters. The environment satisfaction lying at
2.72, indicates that employees have a neutral to positive sentiment towards the
comfort and employees within their workplace.
Patterns of career advancement are quite different. The average hourly pay is
$65.98 (SD = 20.21), ranging from $30 to $100 per hour. Salary hikes average
15.22%, with increments from 11% to 25%. Employees have an average of industry
experience of 11.33 years but only 7.08 years with the current company. Role change
occurs around every 4.29 years, but some workers go up to 15 years without a
promotion, suggesting a lack of career growth for some.
On the 2 to 4 scale of performance ratings, the mean is 2.95 with SD 0.52
suggesting that most employees have comparable evaluations which is quite stable.
Finally, training participation is moderate because attendance at training sessions
per employee averages 2.79 per year with a range of 0 to 6 times last year, which may
limit growth and development in a career.
The visualization of data in the analysis highlights some key aspects of the
employees, their level of satisfaction, the compensation, and the career progression.
Connecting these aspects will further enrich our understanding of what drives job
satisfaction, retention, and performance.
 CHAPTER 2: EXPLORATORY DATA ANALYSIS (EDA)
2.1 Univariate Analysis
2.1.1 Categorical Variables
Figure 2.1. Distribution of categorical variables

Source: The authors (2025)

Based on the gender distribution, it is apparent that there are more male than
female employees which may have some effects on workplace relations and diversity
policies.
The majority of employees not working overtime indicates that for most people,
there is some level of work-life balance. Nonetheless, a significant portion of
employees do work overtime, which may have several implications for their job
satisfaction and productivity.
Employees' attrition rates imply that most of them are staying with the company,
although some have left. The causes of employee turnover or retention due to job
discontent, advancement possibilities, or a heavy workload could provide useful
guidance on retention plans.
The most prominent married group is followed by single and divorced groups.
This may have consequences associated with work-life balance as well as policies
and benefits catered for different life stages.
 The majority of employees travel infrequently while a small proportion travel
frequently. Employees who travel frequently may have higher demands of work
requiring them to be more engaged or may be more stressed.
Analysis of academic backgrounds show that a large number of employees have
qualifications from Life Sciences and Medical fields while less employees have
qualifications in Marketing, Technical areas, or Human Resources. This means that
employees, to a large extent, specialize in scientific and technical areas.
In terms of the distribution among the departments, greater employment work in
Sales, Development and Research and Development and less employment work in
Human Resource and Finance department. It seems that this distribution corresponds
to the company priorities and its main directions of work.
Lastly, job roles vary greatly, with a concentration in executive and technical
jobs, compared to more specialized positions with less employees. This pattern of
employment suggests that the organization emphasizes research, sales, and operation.
These categorized observations help analyze the workforce structure, potential
organizational problems, and identify possibilities for improving employee
motivation and retention.
2.1.2. Numerical Variables
According to the Performance Rating chart, most employees are rated 3.
Ratings of 2 and 4 are outliers, while none of the employees are rated 1 or 5, which
indicates a narrowed evaluation scale.
Employee Environment Satisfaction is mostly rated 3 or 4, suggesting general
satisfaction. The box plot suggests a balanced distribution with no major outliers.
Employee Last Salary Hike Percent range between the 11% and 15%, with a
gradual decline in higher hikes, creating left-skewed distribution with some outliers at
the upper end.
The Employee Work-Life Balance is moderate (average = 3), showing that
most employees are satisfied with their work life balance. The box plot indicates a
concentrated distribution with no significant outliers.
Experience Years in Current Role right-skewed, showing that most employees
have 5 years or less, The figure decreases with greater experience. There are also
outliers as some employees have remained in the same role longer than 15 years.
 Figure 2.2. Distribution and boxplots of selected numerical variables

Source: The authors (2025)

Years with Current Manager chart shows frequent managerial changes, with a
very small number of employees having the same manager for over 15 years. There
are some outliers as many employees have remained under the same leadership for an
unusually long time.
The Years Since Last Promotion chart demonstrates that most employees were
promoted within the last 2 years. The likelihood of promotion is higher in initial
years. The distribution is right-skewed, with many outliers beyond ten years,
highlighting career stagnation for some.
Experience Years at The Company follows a highly right-skewed pattern, with
most employees having less than 10 years, with a sharp decline indicating fewer
long-tenured employees. The box plot highlights several outliers with some
employees having up to 40 years of experience.
Employee Job Satisfaction remains fairly constant, with a slight preference in
higher levels (3 and 4), with no significant outliers.
 Employee Job Involvement is mostly moderate (many rated as 3), followed by
rating 2. The box plot confirms the spread of responses is stable as no extreme
outliers exist.
Employee Relationship Satisfaction presents a fairly even distribution, with
satisfaction ratings of 3 and 4 being more common, without significant outliers.
Employee Job Level is right-skewed, with most workers in the junior levels
(level 1 & 2) and fewer in senior rank (level 3 & 4). No extreme outliers are apparent.
The Number of Companies Worked chart illustrates that many employees
have worked at a single company, with the declining trend as previous company
count rises. The box plot features one outlier with extensive job-hopping.
The Training Times Last Year chart shows that most had either 2 or 3 training
sessions, with very few having 0 and some having 5 or 6 as outliers.
The Employee Hourly Rate chart demonstrates a uniform distribution with no
extreme values, which means that income levels are spread across the scale.
Employee Education Level is evenly distributed, with levels 3 and 4 being the
most common, while fewer have levels 1 or 5. The box plot suggests no clear outliers.
2.2 Bivariate & Multivariate Analysis
2.2.1 Overview of Employee Performance Rating
Table 2.1: Distribution of employee performance ratings
PerformanceRating Count
3 874
2 194
4 132
Source: The authors (2025)
An analysis reveals a distribution of performance rating using the
.value_counts() function, with Rating 3 being the most common (874 employees),
making up 72.8% of the dataset. Meanwhile, the figure for Rating 2 and Rating 4 is
much lower, accounting only 194 and 132 employees, respectively. The significant
gap observed from the given dataset reveals that there are further potential insights.
Next, the authors categorized employees based on gender and age. First, by
determining the minimum and maximum ages, our group segmented the dataset into
five age groups as the provided bar chart, using pd.cut() function, allowing for a
structured comparison that covers different age categories. Subsequently, we grouped
the given dataset based on gender and performance rating and used .unstack() for
counting occurrences. Similar approach was also applied to the age group. Missing
values (NaN) were replaced with 0 by using .fillna(0) to ensure comprehensiveness.
Finally, the group of authors visualized the result by a pie chart and bar chart, putting
them together for easier comparison.
Figure 2.3: Overview of employee performance ratings

Source: The authors (2025)

Overall, the pie chart reveals that the distribution of employee performance
ratings shows a strong concentration at Rating 3, accounting for 72.8% of the
dataset, suggesting deeper insights into the efficiency of the organization’s evaluation
system. When the system rates nearly three quarters of employees the same rank, it
means that the organization fails to adequately distinguish between different levels.
However, the tendency to give employees mid-range levels to avoid conflicts by
some managers may weaken motivation for higher performers to contribute to the
work because of being undervalued.
Breaking down the result by gender, we found that the number of male and
female employees receiving Rating 3 witnesses a huge discrepancy, with the former
comprising 525 people and the latter comprising only 349 people. Furthermore, as
illustrated in the last chart, the figure for employees aged 25-34 receiving Rating 3
surpasses the other age groups. It can be explained in several ways. During such an
age range (25-44), the employees have accumulated sufficient experience to perform
steadily. In contrast, the younger age group (<25) shows a lower performance,
possibly due to being in the adaptation stage. Similar trend is also true for older
employees (55+), which is possibly attributed to the adaptation to technological
changes, physical and cognitive demands or different evaluation criterias. ​
2.2.2 Employee Performance by Department & Job Involvement
To better understand and uncover insights from employee performance data, the
group of authors first calculated the total of performance ratings for each department.
To maintain the chart’s clarity and simplicity, we categorized all other departments
under “Others”, visualizing the final results on the pie chart below. Taking one step
further, the authors also examined the average performance rating of employees
within departments in a gender-based approach with a view to conducting a more
comprehensive analysis.
Figure 2.4: Department-wise employee performance

Source: The authors (2025)

From the given pie chart, it is evident that the Development department holds
the highest performance rating, accounting for 31.5% of the company’s overall
productivity. Following closely behind are the Sales (30.2%) and Research &
Development (28.3%) department. Collectively, these three divisions take up nearly
90% of the total employee performance ratings, highlighting their strategic
importance within the organization and their critical role in driving the company’s
success. The dominance of these three departments can be attributed to more
performance-based incentives and the nature of work, which is closely tied to revenue
generation, product innovation, and relationship building.
The provided stacked bar chart illustrates the average gender-based performance
rating across various departments. However, no significant gap between the female
and male performance rating is evident, notwithstanding minor variations. The
Development department, despite having the highest total performance rating,
maintains a relatively balanced between genders. This finding suggests that gender
does not play a decisive role in evaluating the performance rating within the
divisions. Any minor disparity observed can be due to other variables such as work
experience, department-specific demands rather than gender-based discrepancy. From
the chart we can clearly observe that the largest gap is in the Finance division, in
which the performance rating of both female and male is lower than the others, at
2.68 and 2.85 respectively, resulting in a disparity of 0.17. Despite small fluctuations,
having the lowest performance rating indicates that further investigation should be
conducted to analyze the reason behind this lower rate to take timely actions.
Figure 2.5: Employee performance across departments, by job involvement

Source: The authors (2025)

Job involvement is considered to be an important intrinsic factor contributing to
employee performance. It can be defined as the degree to which employees
“emotionally and cognitively” invest in their work (Hngoi et al., 2024) with a
perception that it occupies a major position in their life interest (Dubin, 1956).
Similar view was also supported by Lodahl & Kejner (1965), Signh & Gupta (2015)
and Salessi & Omar (2019).
In the provided chart, an obviously contradictory trend can be observed.
Generally, as the job involvement rating within departments increases, performance
rating experiences a downward trend, highlighting that higher engagement levels do
not necessarily result in higher productivity. Greenhalgh and Rosenblatt (1984)
indicated that the heavy workload potentially causes anxiety for employees and
drives them to distraction, which negatively impacts individual performance (called
“productivity paradox”). However, there is limited research on the relationship
between job involvement and employee performance, or if yes, they produced
divergent results, suggesting that the relationship between these two variables may
vary depending on contextual factors.
The sales department stands out as the only department in the chart witnessing
an upward trend in employee performance since the job involvement increases. The
driving force behind this increase can be due to a widely held perception. It is usually
perceived among people working in the Sales department that higher involvement
would directly lead to higher sales, more client acquisition and more deal closures.
Another reason can be due to the nature of work in this department, which follows a
commission-based structure that serves as a strong motivator. Generally, a higher
level of involvement could result in more financial rewards, making the positive
correlation between involvement and performance become more apparent.
2.2.3 Employee Performance & Workplace Factors: Job Level, Work-Life Balance
& Job Satisfaction
Figure 2.6: Employee performance affected by workplace factors

Source: The authors (2025)

After assessing how job engagement influences the performance within an
organization, the authors also examine other factors regarding the workplace. These
factors include job level, work-life balance, and job satisfaction. These correlations
are adequately demonstrated by the writers in three different above bar charts.
Overall, it is evident that job level has a negative correlation with average
performance rating while work-life balance and environment satisfaction witnesses a
gradual and significant increase, suggesting that these workplace-based determinants
have various degrees of impacts on their individual performance.
 To be more specific, observing from the first bar chart that there is a slight
decrease in the average rating as the job level increases. In other words, moving
up to a higher job level does not directly translate into higher performance. In
contrast, employees experienced a reduction in their productivity. Holding a higher
job role means that they may deal with extreme job demands or role overload, which
could lead to emotional exhaustion. Past researchers also proved that role overload
has a strong impact on employee’s mental health, such as Janssen et al., (1999),
which could potentially result in lower performance rating at work. The general trend
of job role on performance across department is clearly illustrated as below:
Figure 2.7 : Employee performance across departments, by job levels

Source: The authors (2025)

The opposite trend is true for the two other factors, which are work life balance
(WLB) and environmental satisfaction. The two bar chart suggests that there is a
positive correlation between two mentioned factors and employee performance. Such
a relationship has attracted considerable attention from researchers in the past.
To be specific, according to Scholarios and Marks (2004), WLB plays an
important role in forming employee attitudes toward an organization, such as
organizational attachment, job satisfaction. In another research, Kanwar et al., (2009)
uncovered that there is also a closely positive relationship between job satisfaction
and WLB. Perry-Smith et al., (2000) suggested that effort of an organization in
adopting more work-family policies is associated with a higher performance.
Moreover, family is of great importance in terms of uplifting performance and
boosting positive energy at work (Baral, 2009). From the data, we also found a
similar trend with these mentioned results, that is employees with higher WLB have a
tendency to perform better at their workplace. This finding suggests that
organizations should implement flexible work schedules and mental health support
programmes to obtain more firm-level achievements.
2.2.4 Employee Performance & Career Growth, Workload: Promotions, Salary
Hikes & Overtime
Figure 2.8: Employee performance affected by promotions & salary hikes

Source: The authors (2025)

Regarding the left line chart, a general downward trend indicates that employees
tend to experience a decline in performance over time when promotions are delayed.
According to Appendix 3, the lowest performance rating of 2.545, occurring at 8
years since the last promotion, may suggest a period of burnout, dissatisfaction, or
disengagement due to an extended lack of career advancement. In contrast, the
highest performance rating of 3.125 is observed at 13 years, indicating that long-term
employees develop expertise which sustains their performance and might push harder
in hopes of a late-stage career boost.
Employees who were recently promoted exhibit relatively higher ratings,
yielding around 3.0. However, there is a notable drop between 0 and 2 years
post-promotion, implying that the initial motivation fades relatively quickly. From
year 2 onwards, performance ratings remain steady but at a lower level than
immediately after a promotion.
Regarding the right stacked area chart and Appendix 4, for salary hikes below
20%, performance ratings remain relatively stable at approximately 3.0 for both
genders. This suggests that salary increases within this range are likely determined by
factors such as tenure, or job role adjustments, rather than individual performance.
 ​ However, a significant shift occurs once the salary hike surpasses 20%.
Between 20% and 23%, male employees exhibit slightly higher average performance
ratings than their female counterparts. Notably, at 24%, female employees reach their
peak performance rating of 4.00, whereas male employees peak slightly lower at
3.40. This could indicate that women receiving this salary hike level presumably
demonstrate exceptional performance to be rewarded similarly to men.
Beyond 24%, performance ratings decline for both genders, dropping to 3.375
for females and 3.20 for males. This implies that, beyond a certain salary hike
threshold, the increasing pressure to maintain high performance may outweigh the
benefits of the raise, leading to diminished productivity.
Figure 2.9: Employee performance across departments, by overtime

Source: The authors (2025)

Overall, while employees in Data Science, Human Resources, Research &
Development, and Sales who work overtime generally exhibit higher performance
ratings than those who do not, employees in the Development and Finance
departments with overtime are associated with slightly lower performance ratings
compared to their counterparts working standard hours.
Notably, Appendix 5 shows that Data Science demonstrates the largest positive
gap of 0.333, suggesting that employees benefit from overtime due that are conducive
to focused and extended work periods. Danesh and Nourdad (2017) determine that
roles requiring creative problem-solving can be enhanced through dedicated time and
techniques, which helps explain why Data Science employees achieve a higher
average performance rating of 3.333 during overtime.
 In contrast, the Finance department illustrates the largest negative gap of -0.144,
indicating that employees in financial roles who put in extra hours may risk burnout
due to repetitive tasks and intense nature of their responsibilities. Sheng et al. (2019)
also found that while moderate time pressure can enhance engagement, excessive
time pressure can impair performance, particularly in high-pressure environments
where there is a shortage of psychological capital and diminished sleep quality. This
helps explain why Finance employees working overtime exhibit the lowest
performance rating of 2.666.
2.3 Correlation Analysis
Figure 2.10: Correlation heatmap (Numerical variables)

Source: The authors (2025)

The correlation matrix computed using the dataframe’s corr() function reveals
several key relationships.
On the positive side, there are strong correlations such as:
EmpJobLevel and TotalWorkExperienceInYears (0.78): Employees with
greater total work experience tend to hold higher job levels, which supports the idea
that accumulated experience drives timely promotions and career advancement.
ExperienceYearsAtThisCompany and ExperienceYearsInCurrentRole
(0.76)/ YearsWithCurrManager (0.76): Employees who have been with the
company longer tend to remain in their current roles and have longer working
relationships with their current managers, suggesting a stable workforce and low
managerial turnover, although it might also indicate limited career mobility within the
organization.
On the negative side, PerformanceRating shows slight declines in relation to
several factors:
PerformanceRating and YearsSinceLastPromotion (-0.17): As mentioned
above, performance ratings decrease as time since last promotion increases. This
suggests that employees who have not been promoted for a long period may
experience reduced motivation or engagement, potentially leading to performance
stagnation or decline.
PerformanceRating and ExperienceYearsInCurrentRole (-0.15): Employees
who remain in the same role for longer periods tend to have marginally lower
performance ratings, which may indicate that prolonged tenure without advancement
can lead to complacency.
Notably, there is no correlation (0) between EmpJobSatisfaction and
PerformanceRating, suggesting that job satisfaction may be more strongly
influenced by factors such as work-life balance, benefits, or team dynamics rather
than performance outcomes alone.
2.4 Check Skewness & Kurtosis
Inferred from the table, other variables remain relatively stable with
insignificant skewness. Only ExperienceYearsAtThisCompany (4.057959) and
YearsSinceLastPromotion (3.539080) exhibit strong positive skewness, suggesting
that a large portion of employees have comparatively short tenures or were promoted
recently. The skew in ExperienceYearsAtThisCompany may indicate high labor
turnover, while the skew in YearsSinceLastPromotion could illustrate limited career
advancement opportunities for some employees. Consequently, these two variables
are processed using a log transformation in Chapter 3 to reduce skewness and
restore a normal distribution within the allowable range.

Figure 2.11: Skewness and Kurtosis for numerical variables

 Source: The authors (2025)

CHAPTER 3 : DATA PRE-PROCESSING

Before entering into the modelling stage, data preprocessing is a significant step
to convert data into meaningful insights. It includes getting the data ready for analysis
while guaranteeing its accuracy and suitability for the intended use.
3.1 Check Missing Value
Firstly, missing values happen when no data is stored for a particular variable in
an observation (i.e., data entry errors, system malfunctions, or respondents not
answering), affecting the model accuracy and insights. Therefore, checking missing
values is necessary in order to verify whether the dataset is complete and reliable for
analysis, thereby enhancing the model performance and statistical validity.
Using the code “df.isnull().sum()”, the authors get the count of null values in the
column. No missing values are found, claiming all the data valid for next steps.
3.2 Handle Outliers
After identifying missing values, the authors begin to detect and handle outliers.
Outliers are the extreme data that are significantly different from others, which distort
and affect the model prediction.
 Here, the authors use the Interquartile Range (IQR) method to identify outliers
in the dataset. By setting up a “fence” outside of Q1 and Q3, any values falling
beyond 1.5 times the interquartile range (Q1-Q3) are considered outliers.

Table 3.1: The list of outliers in the dataset

Variables IQR Outliers
NumCompaniesWorked 39
TotalWorkExperienceInYears 51
TrainingTimesLastYear 188
ExperienceYearsAtThisCompany 56
ExperienceYearsInCurrentRole 16
YearsSinceLastPromotion 88
YearsWithCurrentManager 11
PerformanceRating 326
Source: The authors (2025)
As such, the authors tackle this situation by the following codes:

A list of numerical columns with outliers is defined as “outlier_cols”. They are

replaced with median values, which are calculated by the code “median_value =
df[col].median()”. This method is proved not to be sensitive to outliers since it does
not remove any, thereby preventing data loss and preserving the overall distribution
(Kwak and Kim, 2017). Meanwhile, the authors exclude
ExperienceYearsAtThisCompany, YearsSinceLastPromotion (skewed data) to
process under log transformation later. Notably, PerformanceRating, the target of
the research, is also not processed in this stage. Unlike features, where extreme values
can mislead a model, outliers in the target often indicate genuine performance
differences that should be considered. Hence, handling the outliers may cause losing
important information about employees who perform exceptionally well or poorly.
3.3 Categorical Data Conversion
When encountering categorical variables which represent qualitative meanings,
it is crucial to convert them into numerical format for machine learning models.
Proper encoding can prevent model bias and enhance predictability.

The authors convert categorical values by Label Encoding. An instance of

LabelEncoder() from sklearn.preprocessing is created. The code loops through each
categorical column in categorical_variables and applies Label Encoding using
fit_transform(). Label Encoding assigns numeric values (0, 1, 2, …) to each unique
category in a column.
3.4 Feature Transformation
Feature transformation is a vital step to adjust distributions and reduce the
skewness of data. Its objective is the model performance optimization through
numerical stability between features.
Specifically, in this dataset, according to the EDA and skewness analysis in
Chapter 2, the skewness of the two variables ExperienceYearsAtThisCompany and
YearsSinceLastPromotion were 1.789055 and 1.974932, respectively. The data
exhibiting highly positively skewed (>1.0) should undergo log transformation, rather
than square root since square root is only applicable for moderate positive skewed
data (0.5 to 1.0) (Higgins, White, & Anzures-Cabrera, 2008). Using the log
transformation for YearsSinceLastPromotion, the codes are executed as below.

In the next step, the authors continue to use the Quartile - Quartile (Q-Q) plots
to compare the distribution of the YearsSinceLastPromotion against a normal
distribution to check normality.

The similar process is applied to ExperienceYearsAtThisCompany, and the

results for the two variables are presented below.
 Figure 3.1: The Q-Q plot for Figure 3.2: The Q-Q plot for
YearsSinceLastPromotion ExperienceYearsAtThisCompany

Source: The authors (2025)

The closer the points are to the diagonal line, the more normally distributed the
data is. As shown in the figures, most points in the middle of the plot nearly align
with the diagonal line, indicating that that log transformation has successfully
mitigated skewness and brought the data closer to normality in general.
3.5 Feature Selection
Feature selection is the final step in our data preprocessing, which determines
valid data that contribute to the target prediction while removing redundant or
irrelevant ones. This aims to improve model interpretability, reduce overfitting, and
enhance computational efficiency by eliminating unnecessary features.
Here, by the following code, the authors first eliminate redundant columns
YearsSinceLastPromotion, ExperienceYearsAtThisCompany (since they have
already been processed through log transformation), and AgeGroup (only added by
the authors for better illustration in the previous part).

The below code identifies and selects features that have at least a 0.1 absolute
correlation with "PerformanceRating" from a correlation matrix (cor_matrix), and
these important features will be used in the models.

Finally, in this stage, the authors figure out the top three variables that most
significantly impact employee performance by analyzing their correlation with
PerformanceRating. First, it extracts correlation values from cor_matrix, removes
PerformanceRating itself to avoid self-correlation, and ranks features based on the
absolute values of their correlations. The strongest correlations are sorted in
descending order. The results indicate that EmpEnvironmentSatisfaction (employee
satisfaction with the work environment), EmpLastSalaryHikePercent (percentage
of the last salary increase), and YearsSinceLastPromotion_logtransform (log
transformed years since the last promotion) are the most influential factors affecting
employee performance.

CHAPTER 4: MACHINE LEARNING MODEL CREATION & EVALUATION

4.1. Data Preparation
The dataset undergoes a well data preparation process to enhance the predictive
performance, focusing on feature selection and train-test splitting.
Feature selection is conducted using correlation analysis, where only features
exhibiting a correlation coefficient of at least 0.1 with the target variable,
'Performance Rating', are retained. As a result, the selected indices include [4, 5, 9,
16, 20, 21, 22, 25, 26] (corresponding to variables: EmpDepartment, EmpJobRole,
EmpEnvironmentSatisfaction, EmpLastSalaryHikePercent, EmpWorkLifeBalance,
ExperienceYearsInCurrentRole, YearsWithCurrManager, and two log-transformed
versions of YearsSinceLastPromotion and ExperienceYearsAtThisCompany)
 Following feature selection, the dataset is split into training and testing sets
using an 80-20 ratio. This reinforces that the training set would be large enough to
capture meaningful patterns while maintaining a separate test set for unbiased model
evaluation. The independent variables (X) include all selected features except
PerformanceRating, which serves as the dependent variable (y). The structured
approach guarantees that the dataset is appropriately prepared to facilitate its
integration into subsequent machine learning model stages.
4.2. Model Selection
Four models have been considered for this study: Support Vector Machine
(SVM), Random Forest, Balanced Random Forest (BRF), and Artificial Neural
Network (MLP Classifier). Each model offers unique advantages suited to the
complexity and characteristics of Employee performance data.
4.2.1. Support Vector Machine (SVM)
SVM was developed in the 1990s by Vladimir N. Vapnik and his colleagues,
specifically used to classify data by finding an optimal line or creating multiple
hyperplanes that maximize the distance between each class in an N-dimensional
space (IBM, 2023). The model is a powerful algorithm applicable for pattern
classification (both linear and nonlinear), regression problems, and outlier detection.
SVM derives a class decision by identifying the closest points in the training
data – called support vectors, then positioning the boundary so that the distance from
these points is maximized. Moreover, by minimizing structural risk (which considers
both training error and the model’s complexity) rather than empirical risk, SVM
could efficiently avoid a potential misclassification of testing data (Hong et al., 2005).
In other words, the larger the margin, the lower the generalization error of the
classifier (Punnoose and Ajit, 2016).
One of the foremost strengths of SVM is its ability to handle high-dimensional
datasets efficiently that fit directly with the employee performance dataset, which
often comprises numerous features – from quantitative metrics such as total work
experience to qualitative variables like education background (Digital Defynd, 2024).
Moreover, employee performance data may be naturally non-linear, with
complex interdependencies among various performance indicators. In such cases,
SVM can also use a technique called the Kernel trick. Whereas, to map the input data
into a higher-dimensional space where the data becomes linearly separable. Thus,
allowing SVM to capture non-linear patterns without incurring computational costs,
making them highly adaptable to diverse performance prediction scenarios
(GeeksforGeeks Organization, 2023).
Some relevance of this model to predict Employee performance lies in similar
reports, such as: “Enhancing Employee Performance Management” by Mourad et al.
(2024), “Prediction of Employee Turnover in Organizations using Machine Learning
Algorithms” by Punnoose and Ajit (2016), or “Unbiased employee performance
evaluation using machine learning” by Nayem and Uddin (2024).
4.2.2. Random Forest (RF)
Random Forest is a widely-used machine learning that combines the output of
multiple decision trees, to reach a single result. Its ease of use has driven its adoption,
as it effectively handles both classification and regression problems (Sruthi, 2024).
The mechanism is to construct a forest of trees based on bagging methods,
where each tree is constructed independently from a randomly selected subset of the
data (Punnoose and Ajit, 2016), then find the best node among random subsets
generated and determine the most important features based on the impurity index
calculated during the generation process (Papineni et al., 2021). Moreover, its ability
to automatically balance data sets is particularly useful when dealing with imbalanced
classes (CFI, 2024) like performance rating metrics.
Similar to SVM, RF performs well with large, high-dimensional datasets,
making it a strong choice for analyzing employee performance data (Alex, 2024).
However, one key difference is that RF counts every row during cross-validation,
thus providing strong training and improving prediction. Yet, the model can suffer
from “overfitting” problems, a common challenge in many of the ensemble bagging
and boosting algorithms, as it always considers the majority-voted target values to
perform the best split at each iteration (Papineni et al., 2021).
 With regard to the analysis of employee performance, some relevant reports can
be listed like “Unbiased employee performance evaluation using machine learning”
by Nayem and Uddin (2024), or “Random forest algorithm for HR data classification
and performance analysis in cloud environments” by Dong (2024).
4.2.3. Balanced Random Forest (BRF)
Machine learning models often face challenges when dealing with imbalanced
datasets, where the classifier favors the majority class while underperforming the
minority class, leading to biased predictions.
BRF, a variation of the RF, is specifically designed to address the class
imbalance problem. Precisely, instead of using the entire imbalanced dataset, BRF
performs undersampling of the majority class so that each tree gets a balanced dataset
between the majority and minority classes (Fadly, 2024). That is also the key benefit
of this model. However, as past researchs show, making class priors equal either by
down-sampling the majority class or over-sampling the minority class is usually more
effective with respect to a given performance measurement, and that down-sampling
seems to have an edge over over-sampling. However, down-sampling the majority
class may result in loss of information, as a large part of the majority class is not used
(Chen, Liaw, and Brieman, 2004).
In contrast, BRF has better accuracy for the minority class, by balancing the
data, it can pay more attention to them, making predictions more accurate. Besides,
BRF also reduced bias compared to the standard models, given that the majority class
can dominate, by giving equal weight to both classes (Fadly, 2024).
4.2.4. Artificial Neural Network - ANN (MLP Classifier)
ANN, particularly the Multi-Layer Perceptron (MLP), is a popular choice for
predicting employee performance due to its proven ability to capture complex,
non-linear relationships inherent in human behavior and organizational dynamics.
MLP, a type of ANN that consists of a class of feedforward artificial neural
networks with an input layer, one or more hidden layers, and an output layer, with
each layer characterized by multiple interconnected nodes, or neurons. These
connections are weighted, and learning occurs by adjusting these weights, thereby
allowing the model to learn complex patterns from input data (Moriarty and R.
Miikkulainen, 1998).
 A key benefit of MLP is the ability to perform hierarchical feature extraction.
As data flows through the hidden layers, the network automatically transforms raw
inputs into increasingly abstract representations, allowing the MLP to capture even
nonlinear patterns and correlations among diverse employee data (Banerjee, 2024).
Another significant advantage of MLP is its adaptability through the process of
backpropagation (shortened for “backward propagation of errors”). Backpropagation
computes gradients of a loss function with respect to the model's parameters and
updates the parameters iteratively to minimize the loss. As a result, the model not
only enhances performance on both training and testing data but also prevents
“overfitting” problems, ensuring that the model remains robust even when the
employee training data is noisy or limited (Jaiswal, 2024).
Studies that support the use of MLP in analyzing Employee performance can be
named: “Identifying employee engagement drivers using multilayer perceptron
classifier and sensitivity analysis” by Nunez-Sánchez et al., (2024), or “Evaluation of
Factors Affecting Employees' Performance Using Artificial Neural Networks
Algorithm: The Case Study of Fajr Jam” by Rahmanidoust and Zheng (2019).
4.3. Model Implementation
4.3.1. Handling Class Imbalance
The presence of class imbalance in the Performance Rating dataset is first
identified through the value counts and performance rating distribution ratio (Table
4.1). The analysis reveals that 72.8% of employees receive a Rating 3, while 16.2%
receive Rating 2, and only 11% receive Rating 4. This imbalance poses a challenge
for model training, as machine learning models tend to favor majority classes, leading
to biased predictions and a potential diminishment in the accuracy of
underrepresented categories.
Table 4.1: Overall Performance Rating Distribution
Performance Rating Count Percentage
3 874 72.8%
2 194 16.2%
4 132 11%
Source: The authors (2025)
 To mitigate this issue, a structured approach is implemented to ensure fair
representation of all classes in the dataset. The class distribution is first analyzed to
quantify the extent of imbalance, followed by the application of appropriate
resampling techniques to enhance predictive performance across all rating levels.
Different strategies are employed depending on the selected model. For the
SVM and RF models, the class_weight = "balanced" parameter is applied to
automatically adjust the weights of each class based on their inverse frequency. This
method ensures that the model gives appropriate importance to minority classes
without excessively favoring the majority class.
For the BRF model, the sampling_strategy = “not majority” is applied to
handle the nature of imbalanced employee performance data more effectively. This
approach resamples all minority classes while keeping the majority class unchanged,
preventing excessive duplication of data. By curbing the influence of the most
frequent category, the model gains a more balanced perspective, reducing the risk of
overfitting while still capturing important patterns in underrepresented groups.
However, for the MLP Classifier, "class_weight" is not inherently available.
Instead, Synthetic Minority Over-sampling Technique (SMOTE) is employed to
handle class imbalance by artificially generating synthetic instances for the minority
classes. Unlike random oversampling, which simply duplicates existing samples and
increases the risk of overfitting, SMOTE effectively enhances minority class
representation by interpolating new data points between existing observations. This
method preserves the overall data distribution while improving the representation of
underrepresented categories, enabling the model to learn more generalized patterns
rather than memorizing repeated instances (Husain et al., 2025). Given that the
dataset contained both numerical and categorical features, SMOTE appears to be
particularly advantageous as it retains the relationships between features while
expanding the training dataset.
By minimizing bias to ensure that the test set remained representative of
real-world scenarios, the model's ability to generalize and deliver fair and reliable
predictions across diverse employee categories has been significantly improved.
Generally, this approach not only enhances the predictive performance across all
rating levels but also contributes to a more equitable evaluation of employees,
reducing the risk of misclassification due to class imbalances.
4.3.2. Model Implementation
4.3.2.1. Support Vector Machine (SVM)

The SVM model constructed a Pipeline, which allows sequential

transformations of the data before feeding it into the final estimator. First,
StandardScaler is utilized to subtract the mean and divide it by the standard
deviation for each feature, which means scaling all features to have approximately
zero mean and unit variance, the model can converge more quickly and avoid bias
toward features with larger numeric ranges. Moreover, the core classifier, SVC is
applied. Within the SVC, several key parameters are presented. The RBF (Radial
Basis Function) kernel projects data into a higher-dimensional space, allowing it to
capture complex, non-linear boundaries between classes. The C parameter serves as
a regularization term, where a higher value forces the model to classify every training
point correctly, possibly at the cost of a simpler decision boundary. The setting of
class_weight = ‘balanced’ is used to deal with imbalance as mentioned above.
Finally, fixing the random seed with random_state = 10 guarantees the
reproducibility of the model.

Once the pipeline is set up, the training process is executed by calling the “fit”
method with the training data and labels. This method first fits the scaler on the
training data, then transforms the data accordingly, and finally trains the normalized
data with the RBF kernel, which allows it to map data into a higher-dimensional
space. For prediction, the pipeline’s prediction method applies the same scaling
transformation, using the parameters learned from the training set, to test the data
before feeding it into the trained SVC to generate class predictions.
4.3.2.2. Random Forest (RF)

Similar to SMV, RF built a Pipeline, with 2 key steps. First, StandardScaler

standardizes the numerical features. Second, model training Random Forest
Classifier, an ensemble learning model that builds multiple decision trees and
aggregates their predictions to enhance accuracy and avoid overfitting. RF Classifier
with parameters, including n_estimators = 200, meaning the model consists of 200
decision trees built independently on different samples of the training data. The
min_samples_leaf = 1 and min_samples_split = 2 parameters, which control the
minimum number of samples required in leaf nodes and for a split, respectively. In
other words, a split continues until the stopping criteria are met. The Gini impurity
criterion, to determine the quality of splits. A random_state = 33 for reproducibility,
and class_weight = “balanced” helping with imbalanced datasets. The n_jobs = -1
setting ensures that all available CPU are utilized for faster computation.

The training process begins with the standardization of the training data by
StandardScaler before passing the data to the Random Forest Classifier, which then
constructs an ensemble of decision trees using the training labels y_train. By
aggregating predictions from multiple trees, the Random Forest classifier reduces
overfitting and improves generalization. Once training is complete, predictions are
made following the same steps as training with the final prediction for each instance
determined by majority voting. In other words, the class that receives the most votes
across all trees is assigned as the ultimate prediction.
4.3.2.3. Balanced Random Forest (BRF)

Like the above models, BFR pipeline also consists of 2 main parts: scaler
(StandardScaler) to standardize features, and model training (BRF Classifier). In
which, BRF Classifier, a variant of the traditional Random Forest Classifier, is
designed to handle class imbalance effectively by undersampling the majority class
before training each tree in the ensemble.
The BRF Classifier in this report is configured with various main parameters
that enhance its ability to handle imbalanced datasets. The parameter
sampling_strategy = ‘not majority’, is to handle class imbalance. The n_estimators
= 200, indicating that 200 decision trees are built along with a controlled max_depth
= 5, restricting the complexity of each tree to prevent overfitting problems. The
replacement = True parameter allows for sampling with replacement, meaning that
individual data points may appear in multiple trees, which in turn enhances diversity.
In contrast to traditional Random Forest implementations, the bootstrap = False
parameter is applied, disabling the typical bootstrapping method. This modification
ensures that each tree is trained on a distinct random subset of the data, further
emphasizing balanced representation across classes. Finally, the random_state = 10
ensures that results are reproducible.

After defining the pipeline, the model is trained by passing the input data
through StandardScaler. This standardized data is then fed into the BRF Classifier,
where an ensemble of decision trees is built based on balanced subsets of the data.
Once training is complete, the model moves to prediction phase, undergoing
the same processing steps as the training. Yet, unlike the traditional Random Forest
Classifier, each tree in the ensemble contributes to the final prediction, and the
majority voting mechanism of the Random Forest plus these individual decisions will
form a single predicted class. As a result, the method not only reduces the variance
but also enhances the overall stability and reliability of the predictions.
4.3.2.4. Artificial Neural Network - ANN (MLP Classifier)

ANN pipeline consolidates multiple steps, namely SMOTE, scaling, and an

MLP Classifier, into a single workflow. The first component of the pipeline is
SMOTE as to handling class imbalance. Similar to above models, a StandardScaler
is applied, ensuring that all features have a zero mean and a one standard deviation.
The final step in the pipeline is the MLP Classifier, a Multi-Layer Perceptron
neural network. The network uses a backpropagation algorithm to adjust the weights
of each neuron in its two hidden layers, with 100 neurons per layer. Other
parameters such as batch_size = 50 and learning_rate_init = 0.01 determine how
the network updates its weights and processes training examples, respectively. A
larger max_iter = 2000 is specified to allow the model sufficient epoch to converge
on a solution. Besides, setting a random_state = 10 ensures that weight initialization
and other random factors remain consistent, thus facilitating the reproducibility.

The training process begins with a pipeline systematically applying each

transformation step in the specified order - SMOTE, scaling, and then fitting to
MLPClassifier - using only the training data. First, SMOTE identifies minority
classes and synthesizes new examples for those classes. Once the data is rebalanced,
StandardScaler standardizes each feature and passes the data to the MLP Classifier.
Once the model has been trained, predictions are generated, go through the
same workflow as the training process, however, SMOTE is only fitted on the
training data and not applied to the test data in practice, meaning no new synthetic
examples are created during this phase. In the end, the processed test data produces
class predictions based on the learned weights and biases.
4.4. Model Evaluation
4.4.1. Support Vector Machine (SVM)
Table 4.2: SVM Classification Report
precision recall f1-score support
2 0.76 0.86 0.81 29
3 0.93 0.90 0.91 184

4 0.63 0.70 0.67 27

SVM Accuracy 0.87 (0.8708) 240
Macro avg 0.77 0.82 0.80 240
Weighted avg 0.88 0.87 0.87 240
Source: The authors (2025)
The model achieves an accuracy of 87%, indicating that it correctly classifies
the majority of samples in the test set. First of all, Rating 2 achieves a precision of
0.76 and a recall of 0.86, indicating some difficulties in capturing this class correctly.
In contrast, Rating 3 demonstrates particularly strong performance, reflected in a
precision of 0.93 and a recall of 0.90, suggesting that the model rarely confuses this
class with others, given the fact that Rating 3 is supported by a substantial portion of
the dataset, which provides SVM with more training data to learn from and improve
its predictions. Lastly, Rating 4 appears to overlap with Rating 3, resulting in a worse
performance within the SVM model with a 0.70 recall and a 0.63 precision.
Observing the SVM Confusion Matrix (Figure 4.1), majority of Rating 2 are
correctly labeled (25), but a few (4) are misclassified as Rating 3. In contrast, Rating
3, having the largest number of samples (165), maintains a high recall; however,
some are mislabeled as either Rating 2 (8) or Rating 4 (11). Last, Rating 4 shows no
disorder with Rating 2 (0) but does exhibit some overlap with Rating 3 (8).
These findings confirm that the model mainly struggles to differentiate between
certain instances of Rating 2 and Rating 3, and to a lesser extent between Rating 3
and Rating 4, while there is no confusion between Rating 2 and 4.
4.4.2. Random Forest (RF)
Table 4.3: Random Forest Classification Report
 precision recall f1-score support
2 0.93 0.93 0.93 29
3 0.95 0.99 0.97 184

4 1.00 0.74 0.85 27

RF Accuracy 0.95 (0.9542) 240

Macro avg 0.96 0.89 0.92 240

Weighted avg 0.96 0.95 0.95 240
Source: The authors (2025)
Random Forest Classifier achieved an overall accuracy of 95.42%, indicating
that the model is highly effective at classifying employee performance datasets. For
each Rating performance, Rating 2, had a balanced precision and recall of 0.93,
meaning it was correctly identified when present and not frequently misclassified as
another class. Rating 3 (with the highest number of samples, 184) had the best recall
score at 0.99, meaning nearly all instances of Rating 3 were correctly identified.
However, Rating 4, had the highest precision (1.00) but a lower recall (0.74)
meaning although the model predicted Rating 4 correctly every time, it failed to
capture all actual instances of this class, thus leading to misclassification.
From observation of the RF Confusion Matrix (Figure 4.1), majority of
misclassifications occurred in Rating 4, where 7 instances were identified as Rating 3,
likely due to feature similarities between the two classes. Additionally, Rating 2 had 2
samples misclassified as Rating 3, but no misclassification into Rating 4.

4.4.3. Balanced Random Forest (BRF)

Table 4.4: Balanced Random Forest Classification Report
precision recall f1-score support
2 0.96 0.90 0.93 29
3 0.95 0.99 0.97 184
4 1.00 0.74 0.85 27
BRF Accuracy 0.95 (0.9542) 240
Macro avg 0.97 0.88 0.92 240
 Weighted avg 0.96 0.95 0.95 240
Source: The authors (2025)
The BRF model demonstrates a strong overall performance, as reflected by its
accuracy of 95.42%, the same result as the traditional Random Forest learning. By
observing, Rating 2 achieves a precision of 0.96 and a recall of 0.90, indicates the
model corrects 96% of the time prediction and identifies 90% of the actual instances
of Rating 2 in the dataset. The performance in Rating 3 is noteworthy, with a
precision of 0.95 and an impressive recall of 0.99, demonstrating an effective capture
of nearly all instances, given that Rating 3 is the majority class with 184 samples.
Rating 4 yields a perfect precision of 1.00 implying that whenever the model predicts
Rating 4, it results in no falsehood. Yet, the recall of 0.74 shows that about ¼ of
actual samples are missed.
Regarding the BRF Confusion Matrix (Figure 4.1), Rating 2, 26 instances
were correctly classified, with only 3 instances misclassified as Rating 3,
demonstrating strong performance. Rating 3 stands out with 183 instances correctly
predicted and only 1 misclassified as Rating 2, indicating exceptional accuracy for
this class. In contrast, Rating 4 shows a degree of difficulty for the model, with 20
instances correctly identified and 7 instances misclassified as Rating 3. These
misclassifications highlight that while the model performs well overall, it tends to
confuse Rating 4 and 3, suggesting areas where further tuning might be needed.

4.4.4. Artificial Neural Network - ANN (MLP Classifier)

Table 4.5: MLP Classification Report
precision recall f1-score support
2 0.85 0.79 0.82 29
3 0.94 0.94 0.94 184
4 0.66 0.70 0.68 27
MLP Accuracy 0.90 (0.8958) 240
Macro avg 0.82 0.81 0.81 240
 Weighted avg 0.90 0.90 0.90 240
Source: The authors (2025)
Overall, the model achieves an accuracy of roughly 0.90 (or 89.58%),
suggesting that it makes correct predictions on the majority of the test samples. In
detail, Rating 3, has the highest support (184 samples), achieves the strongest
performance, yielding a precision and recall of 0.94. Ratings 2 and 4, by contrast,
have far fewer samples (29 and 27, respectively). While Rating 2 still demonstrates
relatively strong metrics, 0.85 precision and 0.79 recall, Rating 4 shows noticeably
weaker performance, with a precision of 0.66, and a recall of 0.70. This pattern may
indicate some confusion between Ratings 3 and 4, as the model appears to struggle in
distinctly identifying Rating 4 samples.
Focusing on Rating 2 for the ANN Confusion Matrix (Figure 4.1), 23
instances are correctly identified, while 3 are incorrectly classified as Rating 3 and
another 3 as Rating 4. Such confusion can arise from overlapping feature
distributions or insufficient representation in the training process. In contrast, Rating
3 correctly classifies 173 instances. Still, there are 4 instances incorrectly labeled as
Rating 2 and 7 incorrectly labeled as Rating 4. Lastly, Rating 4 with 19 instances
accurately classified, yet, 8 instances were misclassified as Rating 3. This pattern is
often observed in situations where Rating 4 is less frequent in the dataset, limiting the
model to distinct patterns and leading to confusion with the more common Rating 3.

Figure 4.1: Summary of the 4 models’ Confusion Matrix

 Source: The authors (2025)

4.5. Model Comparison

Figure 4.2: Accuracy comparison of 4 models

Source: The authors (2025)

 In evaluating the employee performance, both RF and BRF emerged as the
top-performing algorithms with an accuracy of 95.42%. These models
outperformed MLP/ ANN (89.58% accuracy), and the SVM (87.08%). From a
high-level perspective, the RF family’s success stems from its ensemble approach of
combining multiple decision trees, leading to strong prediction and preventing
overfitting. This makes it particularly suitable for complex datasets where employee
performance is influenced by both categorical and numerical variables.
Delving into the classification reports, RF and BRF not only produce high
overall accuracy but also exceed in class-level metrics with strong precision and
recall scores across the different classes (Ratings 2, 3, and 4). In a business context, it
implies that the model is reliable for identifying both high and low performers. In an
employee performance scenario, where “excellent” or “low” performers might be
underrepresented, the BRF model ensures that these classes receive due attention,
leading to more equitable predictions.
From a human resources perspective, these findings underscore the importance
of using ensemble methods to gain nuanced insights into workforce productivity. The
high recall for certain classes (notably Rating 3 in both RF family models) means that
the model captures a large proportion of employees truly belonging to that
performance level. High precision, on the other hand, ensures that when the model
predicts a certain performance level, it is correct most of the time. This balance of
precision and recall translates into more confident decision-making, whether the goal
is to identify high-potential employees for leadership programs or to detect
underperformers who may benefit from targeted interventions and support.
 CHAPTER 5: KEYS FINDINGS & RECOMMENDATIONS
5.1. Key Findings
The above evaluation of employee performance data provided key insights into
the factors impacting productivity and engagement. The authors used statistical
correlations and machine learning models to identify key variables affecting
performance ratings, and the top 3 variables with the highest correlation to
performance are EmpEnvironmentSatisfaction, EmpLastSalaryHikePercent, and
YearsSinceLastPromotion. These variables will be reviewed to summarize their
significance on the overall employee performance.
5.1.1. Employee Environment Satisfaction
Employee environment satisfaction has the strongest correlation with
performance ratings (correlation coefficient equals 0.4), emphasizing the importance
of a positive work environment in driving employee engagement and overall
productivity. This result is reasonable as a conducive work environment provides
employees with a sense of security, thus enabling them to perform optimally
(Badrianto et al., 2020). Moreover, the working environment can affect employee’s
emotions. Therefore, when the employees are satisfied with their working
environment, they tend to feel comfortable, which allows them to perform their tasks
efficiently. As a result, they make optimal use of their working hours, leading to
increased productivity and enhanced performance (Ramli et al., 2019).
5.1.2. Last Salary Hike Percentage
Employees who have recently received salary increases also demonstrate
significant improvements in their work performance levels (correlation coefficient
equals 0.33). This result suggests that financial incentives act as a powerful working
motivator, emphasizing the importance of maintaining a competitive compensation
structure to enhance employee motivation and retention. Employees who receive a
sufficient salary to cover their daily expenses are more likely to work efficiently,
particularly when their contributions are acknowledged through bonuses, incentives
and salary increases (AI Mehrzi et al,. 2016).
5.1.3. Years Since Last Promotion
The result also found that employees who receive recent promotions are more
likely to demonstrate higher working performance, with a correlation coefficient
equal to - 0.24 (after categorical encoding and log transformation), showing that
career promotion opportunities enhance employee motivation and engagement. Job
promotions offer employees opportunities for personal development, greater
responsibilities and power, and also enhanced social status. The desire for a
promotion motivates high-performing employees to put in extra effort to reach their
performance goals (Nguyen et al., 2015). When promotions are effectively
implemented, they contribute to increased job satisfaction, thus improving employee
performance (Robbins and Judge, 2011). Especially in high power distance cultures,
people may value promotions more than bonuses, as promotions signify higher social
status and success (Nguyen et al., 2015).
5.2. Strategies to Improve Employee Performance
To ensure a positive work environment, companies should conduct periodic
employee satisfaction surveys and try to improve based on that feedback. For
example, Microsoft has actively gathered feedback from employee engagement
surveys and adjusted accordingly to improve employee satisfaction and the
organization’s overall health (Microsoft, 2018). Moreover, it is necessary to
encourage open communication and foster a culture of inclusivity and diversity,
which can help to build a more positive and collaborative atmosphere. Google, for
instance, has successfully implemented open office designs, flexible working
arrangements, and intentionally flat organizational structures to enhance creativity
and job satisfaction, which also serve as an attractive feature for top talent (Abirami,
n.d.). Finally, prioritizing ergonomic workspace design, cutting-edge technology, and
comprehensive wellness programs also significantly contribute to productivity, and
enhance employee satisfaction and overall well-being.
In terms of salary and bonuses, a well-defined performance-based salary
increment framework should be established to maintain workforce motivation.
Beyond salary increases and financial incentives, companies can also offer
non-monetary benefits such as work-from-home days, professional development
opportunities, employee recognition programs, or travel packages. It is also crucial
for companies to regularly conduct industrial benchmarks to ensure a competitive and
reasonable salary structure, adding to talent retention and motivation. By addressing
both financial and non-financial needs, firms can create a more balanced and
supportive working environment, which will ultimately drive higher productivity,
innovation, and employee loyalty.
Regarding career development, transparent career progression pathways and
promotion criteria should be implemented to provide employees with clear growth
opportunities. Leadership development programs can play a crucial role in preparing
employees for higher position responsibilities, thereby ensuring a steady flow of
skilled employees ready to take on further career ladder. Moreover, companies should
also conduct cross-functional mobility and promote job diversification through
rotation programs to broaden staff’s competencies. For instance, companies such as
Unilever or Nestle have successfully implemented rotational leadership programs
with clear career paths to equip employees with hands-on experience across various
functions and accelerate their career progressions.

CONCLUSION
In conclusion, the authors have identified the most influential factors impacting
employee performance, including employee environment satisfaction, last salary hike
percentage, and years since last promotions. Several strategies to enhance workplace
performance ratings and productivity have also been suggested. Specifically, by
fostering a positive working environment, implementing performance-based salary
systems, and providing transparent career development opportunities for employees,
companies can create a more motivated, innovative and loyal workforce. While
financial incentives remain important, non-monetary benefits, suitable organizational
structure and job promotions also significantly enhance employee efficiency.
Regarding limitations, the constraint comes from the dataset itself, which
contains only 1200 records. This sample size is relatively small, thus potentially
failing to fully reflect the diversity and complexity of a broader workplace. Moreover,
this dataset is missing some other important details that could also affect employee
performance, such as the actual amount of salary (since in the dataset, it only includes
percentage raises), working conditions like hours or workload, and some outside
factors like company culture or organizational structure. The absent information
limits the depth of the analysis and its ability to comprehensively explain
performance variations among employees.
 REFERENCE

Abirami, D. M. Open Office Design and its Impact on Employees: A Review of

Research and Perspectives.
Alex, S. (2024, September 30). What are the Advantages and Disadvantages of
Random Forest? Retrieved from Pickl.AI website:
https://www.pickl.ai/blog/advantages-and-disadvantages-random-forest/
Al Mehrzi, N., & Singh, S. K. (2016). Competing through employee engagement: a
proposed framework. International Journal of Productivity and Performance
Management, 65(6), 831-843.
Badrianto, Y., & Ekhsan, M. (2020). Effect of work environment and job satisfaction
on employee performance in pt. Nesinak industries. Journal of Business,
Management, & Accounting, 2(1).
Banerjee, S. (2024, April 6). Exploring the Power and Limitations of Multi-Layer
Perceptron (MLP) in Machine Learning. Retrieved from Medium website:
https://shekhar-banerjee96.medium.com/exploring-the-power-and-limitations-
of-multi-layer-perceptron-mlp-in-machine-learning-d97a3f84f9f4
Baral, R. (2009). Examining antecedents of work-family enrichment and its effect on
individual, family and organisational outcomes. Unpublished Doctoral
Dissertation, IIT Bombay.
Chen, C., Liaw, A., & Brieman, L. (2004, January). Using Random Forest to Learn
Imbalanced Data | Department of Statistics. Retrieved from
statistics.berkeley.edu website: https://statistics.berkeley.edu/tech-reports/666
Dahkoul, Z. M. (2018). The determinants of employee performance in Jordanian
organizations. Pressacademia, 5(1), 11–17.
https://doi.org/10.17261/pressacademia.2018.780
Danesh, M., & Nourdad, N. (2017). On the Relationship between Creative Problem
Solving Skill and EFL Reading Comprehension Ability. Theory and Practice
in Language Studies, 7(3), 234. https://doi.org/10.17507/tpls.0703.10
Digital Defynd. (2024, July 6). 10 Pros & Cons of Support Vector Machines [2024].
Retrieved from DigitalDefynd website:
https://digitaldefynd.com/IQ/pros-cons-of-support-vector-machines/
Dong, F. (2024). Random Forest Algorithm for HR Data Classification and
Performance Analysis in Cloud Environments. International Journal of
Advanced Computer Science and Applications, 15(11).
https://doi.org/10.14569/ijacsa.2024.0151147
Dubin, R. (1956). Industrial Workers’ Worlds: A Study of the “Central Life Interests”
of Industrial Workers. Social Problems, 3(3), 131–142.
https://doi.org/10.2307/799133
Fadly, M. T. (2024, July 3). BALANCED RANDOM FOREST - Fadly Mochammad
Taufiq - Medium. Retrieved March 22, 2025, from Medium website:
https://medium.com/@fadleemt/balanced-random-forest-d5dc9c896bb4
GeeksforGeeks Organization. (2023, May 7). Support Vector Machine in Machine
Learning. Retrieved from GeeksforGeeks website:
https://www.geeksforgeeks.org/support-vector-machine-in-machine-learning/
Greenhalgh, L., & Rosenblatt, Z. (1984). Job insecurity: Toward conceptual clarity.
Academy of Management review, 9(3), 438-448.
Higgins, J. P. T., White, I. R., & Anzures-Cabrera, J. (2008). Meta-analysis of skewed
data: Combining results reported on log-transformed or raw scales. Statistics
in Medicine, 27(29), 6072–6092. https://doi.org/10.1002/sim.3427
Hngoi, C. L., Abdullah, N. A., Wan Sulaiman, W. S., & Zaiedy Nor, N. I. (2024).
Examining job involvement and perceived organizational support toward
organizational commitment: job insecurity as mediator. Frontiers in
Psychology, 15, 1290122.
Hong, W.-C., Pai, P.-F., Huang, Y.-Y., & Yang, S.-L. (2005). Application of Support
Vector Machines in Predicting Employee Turnover Based on Job Performance.
Lecture Notes in Computer Science, 3610(978-3-540-31853-8), 668–674.
https://doi.org/10.1007/11539087_85
IBM. (2023, December 12). Support Vector Machine. Retrieved from IBM website:
https://www.ibm.com/think/topics/support-vector-machine
Jaiswal, S. (2024, February 7). Multilayer Perceptrons in Machine Learning: A
Comprehensive Guide. Retrieved from Datacamp.com website:
https://www.datacamp.com/tutorial/multilayer-perceptrons-in-machine-learnin
g
Janssen, P. P., Schaufelioe, W. B., & Houkes, I. (1999). Work-related and individual
determinants of the three burnout dimensions. Work & stress, 13(1), 74-86.
Kanwar, Y. P. S., Singh, A. K., & Kodwani, A. D. (2009). Work—life balance and
burnout as predictors of job satisfaction in the IT-ITES industry. Vision, 13(2),
1-12.
Kwak, S. K., & Kim, J. H. (2017). Statistical data preparation: management of missing
values and outliers. Korean Journal of Anesthesiology, 70(4), 407–411.
https://doi.org/10.4097/kjae.2017.70.4.407
Lasa, A. N., Pedroni, A., & Komm, A. (2024, May 15). Performance management
that puts people first. Retrieved from www.mckinsey.com website:
https://www.mckinsey.com/capabilities/people-and-organizational-performanc
e/our-insights/in-the-spotlight-performance-management-that-puts-people-first
Lodahl, T. M., & Kejnar, M. (1965). The definition and measurement of job
involvement. Journal of applied psychology, 49(1), 24.
Mesiya, A. Y. (2019, June 30). Factors Affecting Employee Performance: An
Investigation on the Private School Sector. Retrieved from papers.ssrn.com
website: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3645036
Microsoft. (2018). Improving employee engagement in the workplace. Retrieved
from Microsoft.com website:
https://www.microsoft.com/en-us/microsoft-365/business-insights-ideas/resour
ces/employee-engagement
Moriarty, D. E., & R. Miikkulainen. (1998). Hierarchical evolution of neural
networks. World Congress on Computational Intelligence, 98TH8360(pp.
428-433). https://doi.org/10.1109/icec.1998.699793
Mourad, Z., Noura, A., Mohamed, C., & Abdelhamid, B. (2024). Enhancing
Employee Performance Management. International Journal of Advanced
Computer Science and Applications/International Journal of Advanced
Computer Science & Applications, 15(3).
https://doi.org/10.14569/ijacsa.2024.01503100
Nayem, Z., & Uddin, A. (2024). Unbiased Employee Performance Evaluation Using
Machine Learning. Journal of Open Innovation: Technology, Market, and
Complexity, 10(1), 100243. https://doi.org/10.1016/j.joitmc.2024.100243
Nguyen, P. D., Dang, C. X., & Nguyen, L. D. (2015). Would better earning, work
environment, and promotion opportunities increase employee performance?
An investigation in state and other sectors in Vietnam. Public Organization
Review, 15, 565-579.
Núñez-Sánchez, J. M., Jesús Molina-Gómez, Pere Mercadé-Melé, &
Fernández-Miguélez, S. M. (2024). Identifying employee engagement drivers
using multilayer perceptron classifier and sensitivity analysis. Eurasian
Economic Review, 14. https://doi.org/10.1007/s40821-024-00283-6
Papineni, S. lakshmi, Reddy, M., Yarlagadda, S., Yarlagadda, S., & Akkineni, H.
(2021). An Extensive Analytical Approach on Human Resources using
Random Forest Algorithm. International Journal of Engineering Trends and
Technology, 69(5), 119–127.
https://doi.org/10.14445/22315381/ijett-v69i5p217
Perry-Smith, J. E., & Blum, T. C. (2000). Work-family human resource bundles and
perceived organizational performance. Academy of management Journal,
43(6), 1107-1117.
Punnoose, R., & Ajit, P. (2016). Prediction of Employee Turnover in Organizations
using Machine Learning Algorithms. International Journal of Advanced
Research in Artificial Intelligence, 5(9).
https://doi.org/10.14569/ijarai.2016.050904
Rahmanidoust, M., & Zheng, J. (2019). Evaluation of Factors Affecting Employees’
Performance Using Artificial Neural Networks Algorithm: The Case Study of
Fajr Jam. International Business Research, 12(10), 86.
https://doi.org/10.5539/ibr.v12n10p86
Ramli, A. H. (2019). Work environment, job satisfaction and employee performance
in health services. Business and Entrepreneurial Review, 19(1), 29-42.
Reginaldo G. de S. Neto, Jatobá, M. N., Santana, M., Fernandes, P. S., Ferreira, J. J.,
Foleis, J. H., & Teixeira, J. P. (2025). Human Resources Sptimization with
MultiLayer Perceptron: An Automated Selection Tool. Procedia Computer
Science, 256(1877-0509), 238–245.
https://doi.org/10.1016/j.procs.2025.02.117
Salessi, S. M., & Omar, A. G. (2019). Job involvement in current research: Update
and state of the art.
Schermerhorn Jr, J. R., Osborn, R. N., Uhl-Bien, M., & Hunt, J. G. (2011).
Organizational behavior. john wiley & sons.
Scholarios, D., & Marks, A. (2004). Work‐life balance and the software worker.
Human Resource Management Journal, 14(2), 54-74.
Sheng, X., Wang, Y., Hong, W., Zhu, Z., & Zhang, X. (2019). The curvilinear
relationship between daily time pressure and work engagement: The role of
psychological capital and sleep. International Journal of Stress Management,
26(1), 25–35. https://doi.org/10.1037/str0000085
Singh, A., & Gupta, B. (2015). Job involvement, organizational commitment,
professional commitment, and team commitment: A study of generational
diversity. Benchmarking: An International Journal, 22(6), 1192-1211.
Sulistiani, L., & Faozanudin, M. (2022). Effectiveness Analysis of the Employee
Work Performance Assessment System – A Critical Three-Component
Approach. Expert Journal of Business and Management, 10(2). Retrieved
from https://business.expertjournals.com/23446781-1005/
Zhenjing, G., Chupradit, S., Ku, K. Y., Nassani, A. A., & Haffar, M. (2022). Impact
of employees’ Workplace Environment on employees’ performance: a
multi-mediation Model. Frontiers in Public Health, 10(890400). NCBI.
https://doi.org/10.3389/fpubh.2022.890400
 APPENDICES

PerformanceRating 2 3 4

Gender

Female 75 349 51

Male 119 525 81

AgeGroup

<25 12 54 11

25-34 69 336 47

35-44 66 300 48

45-54 36 151 18

55+ 11 33 8

Appendix 1: Average employee performance ratings by gender and age

df.groupby(['Gender', 'PerformanceRating']).size().unstack().fillna(0)
df.groupby(['AgeGroup', 'PerformanceRating']).size().unstack().fillna(0)

EmDepartment PerformanceRating

0 Development 1114

1 Sales 1067

2 Research & Development 1002

3 Human Resources 158

4 Finance 136

5 Data Science 61

EmDepartment Gender PerformanceRating

0 Data Science Female 3.000000

1 Data Science Male 3.083333

2 Development Female 3.098592

3 Development Male 3.077626

 4 Finance Female 2.681818

5 Finance Male 2.851852

6 Human Resources Female 3.058824

7 Human Resources Male 2.864865

8 Research & Development Female 2.945736

9 Research & Development Male 2.906542

10 Sales Female 2.840764

11 Sales Male 2.875000

Appendix 2: Distribution of employee performance ratings across departments,

by gender
df.groupby('EmpDepartment')['PerformanceRating'].sum()
df.groupby(['EmpDepartment', 'Gender'])['PerformanceRating'].mean()

YearsSinceLastPromotion PerformanceRating

0 0 3.123667

1 1 2.898990

2 2 2.763780

3 3 2.777778

4 4 2.773585

5 5 2.828571

6 6 2.833333

7 7 2.854839

8 8 2.545455

9 9 2.687500

10 10 2.600000

11 11 2.826087

12 12 2.666667

13 13 3.125000
14 14 3.000000

15 15 2.909091

Appendix 3: Average performance rating by years since last promoted

df.groupby(['YearsSinceLastPromotion'])['PerformanceRating'].mean().reset_index()

EmpLastSalaryHikePercent Gender PerformanceRating

0 11 Female 2.828571

1 11 Male 2.848485

2 12 Female 2.846154

3 12 Male 2.800000

4 13 Female 2.793651

5 13 Male 2.895238

6 14 Female 2.843750

7 14 Male 2.870370

8 15 Female 2.928571

9 15 Male 2.907407

10 16 Female 2.812500

11 16 Male 2.888889

12 17 Female 3.000000

13 17 Male 2.869565

14 18 Female 2.928571

15 18 Male 2.822222

16 19 Female 2.892857

17 19 Male 2.857143

18 20 Female 3.440000

19 20 Male 3.280000

20 21 Female 3.454545
21 21 Male 3.652174

22 22 Female 3.444444

23 22 Male 3.413793

24 23 Female 3.363636

25 23 Male 3.700000

26 24 Female 4.000000

27 24 Male 3.400000

28 25 Female 3.375000

29 25 Male 3.200000

Appendix 4: Average performance rating by salary hike and gender

df.groupby(['EmpLastSalaryHikePercent','Gender'])['PerformanceRating'].mean().
reset_index()

EmpDepartment OverTime PerformanceRating

0 Data Science No 3.000000

1 Data Science Yes 3.333333

2 Development No 3.107884

3 Development Yes 3.041667

4 Finance No 2.810811

5 Finance Yes 2.666667

6 Human Resources No 2.850000

7 Human Resources Yes 3.142857

8 Research & Development No 2.894737

9 Research & Development Yes 2.989583

10 Sales No 2.830189

11 Sales Yes 2.935185

Appendix 5: Average performance rating across all departments, by overtime

df.groupby(['EmpDepartment', 'OverTime'])['PerformanceRating'].mean().reset_index()