









































Pa
ge

 
1



Pa
ge

 
16

American Journal of  Smart 
Technology and Solutions (AJSTS)

Addressing Class Imbalance in IoT: A Comparative Analysis of  Resampling Techniques
Yousef  Qawqzeha1*

Volume 4 Issue 1, Year 2025
ISSN: 2837-0295 (Online)

DOI: https://doi.org/10.54536/ajsts.v4i1.2912
https://journals.e-palli.com/home/index.php/ajsts

Article Information ABSTRACT

Received: July 09, 2024
Accepted: August 01, 2024
Published: February 18, 2025

In modern times, automated task processing and sophisticated algorithm design are important 
tools for using cutting-edge technologies and approaches to extract insights from data and 
practical solutions. The machine learning models powered by data have produced outputs 
that were either more or less worthy when the input datasets were balanced. An uneven 
distribution of  classes in the input datasets has resulted in imbalanced data. Class imbalance 
has been a significant challenge in machine learning applications, particularly when working 
with substantially disparate distributions like those found in Internet of  Things datasets. 
This study addressed the class imbalance issue in IoT data by comparing various resampling 
strategies. The study aimed to find efficient ways to realign class distributions and enhance 
the functionality of  machine learning models implemented in Internet of  Things systems. 
A predictive model built on an unbalanced data set appeared to have high accuracy, but it 
struggled to generalise new data from the minority class. Resampling techniques, including 
Over-sampling, Under-sampling, SMOTE (Synthetic Minority Over-Sampling Technique), 
and ADASYN (Adaptive Synthetic Sampling), were evaluated using an extensive variety of  
IoT datasets spanning different classes and domains. The functionality of  each technique was 
assessed using performance metrics such as the area covered by AUC, F1-score, precision, 
and recall. This study advanced the understanding of  class imbalance mitigation in IoT data 
processing by providing insights into creating more durable and trustworthy models for IoT 
scenarios.  CCS CONCEPTS • Class Imbalance • Applied Computing • Machine Learning 
• Internet of  Things (IoT)

Keywords
Multi-Class Classification, 
Resampling Techniques, Class 
Imbalance, Hyperparameter 
Tuning, Fraud Detection

1 University of  Fujairah, 48CP+J4P - E89, Mraisheed, Fujairah, United Arab Emirates
* Corresponding author’s e-mail: YousefQawqzehaa@outlook.com

INTRODUCTION
The Internet of  Things (IoT) has recently changed 
several industries. It has made it possible to collect and 
analyse an immense quantity of  sensor data for various 
applications, from smart healthcare to the automobile 
industry (Pramanik et al., 2019). The Internet of  Things 
(IoT) is a significant advancement in artificial intelligence, 
transforming our daily lives through various functions like 
device modelling, control, data publishing, analysis, and 
detection (Wanasinghe et al., 2020). It has outpaced other 
technologies due to its promising future and ability to 
analyse and study various elements, making it a significant 
milestone in the field (Nord et al., 2019).
However, class imbalance in the datasets has been one 
of  the main obstacles to fully utilising the potential of  
IoT data. When one class greatly outnumbers the others, 
class imbalances arise, which bias model results and lower 
predicted accuracy. If  it is discovered that the amount of  
data points in two-class classification models or multi-
class data models is roughly the same, handling a dataset 
with sufficient data points is not too challenging (Peng et 
al., 2023; Qawqzeh & Ashraf, .2023). 
The utilisation of  IoT generates non-stationary data 
streams that can change over time, making it challenging 
for Machine learning algorithms to identify minority 
exposure accurately (Nixon et al., 2019). The lack of  
robust computing equipment can disrupt machine 
learning methods like oversampling and undersampling, 
affecting their ability to operate in complex environments 

(Atuhurra et al., 2024). Cyberattacks also tax IoT networks, 
leading to highly skewed datasets. Momentum detection 
of  minority-class hacking is crucial in IoT networks, 
but models tend to favour the majority of  normal-class 
sites. Deep learning techniques for class imbalance have 
been applied to image recognition, but their application 
to non-image IoT data may require different approaches 
(Atuhurra et al., 2024; Johnson & Khoshgoftaar, 2019).
This research aims to reduce class disparity in two IoT 
datasets: IoT Modbus and IoT GPS Tracker. Addressing 
class imbalance is crucial for ensuring this reliability and 
effectiveness in machine learning models deployed in IoT 
systems (Qawqzeh & Ashraf, 2023; Tanha et al., .2020; 
Varotto et al., 2021; Welvaars et al., 2023). The main goal 
is a competent and comparative analysis of  methods 
designed to address class imbalance in these datasets. 
In particular, we examine the effectiveness of  several 
techniques, such as the Synthetic Minority Over-Sampling 
Technique (SMOTE), Random Under-sampling (RUS), 
Random Over-sampling (ROS), and Adaptive Synthetic 
Sampling (ADASYN) method, in resolving class imbalance 
in multi-class scenarios. Machine learning algorithms 
trained on imbalance datasheets tend to perform poorly on 
minority class cases, which are generally more interesting 
in detection and forecasting scenarios, in favour of  the 
majority class (Koziarski et al., 2020). 
Multi-class classification is a task with more than two 
classes and assumes that an object can only receive 
one classification. The trained model, constructed with 



Pa
ge

 
17

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

this dataset, will function according to the authors’ 
expectations. It is common knowledge that these data 
points are balanced datasets. 
However, the issue arises when skewed datasets, such as 
those with under- or over-representation, are acquired to 
create a data model for predictive analysis. We seek to 
find practical approaches for enhancing the predictive 
accuracy and generalisation of  machine learning models 
in the Internet of  Things applications by assessing these 
methods’ effectiveness on various IoT datasets. 
This ‘study’s goal was to advance state-of-the-art IoT data 
analysis and make it easier to create more durable and 
trustworthy predictive models for practical IoT scenarios 
by offering insights into the selection and application of  
resampling techniques designed to address class imbalance 
in IoT datasets (Obaid & Nassif, .2022; Paisitkriangkrai et 
al., 2013; Wang & Yao, .2012). 

LITERATURE REVIEW
Introduction to Internet of  Things
The Internet of  Things (IoT) is a rapidly evolving 
technology that uses processing power, downsized 
electronics, and networking links to connect devices 
and systems (Kumar et al., 2021). It has sparked 
debates on various aspects, including opportunities 
for new companies, security, privacy, compatibility, and 
international ecosystem. The IoT will impact various 
aspects of  life, including elders, consumers, and healthcare 
providers (Pal et al., 2018). To ensure energy savings and 
elasticity, IoT devices like smart home appliances need 
authentication and optimization of  energy consumption. 
This technology has the potential to revolutionize various 
aspects of  our lives (Powroźnik et al., 2021). Personal IoT 

devices, such as wearable fitness and health monitoring 
devices, are also expected to improve independence and 
quality of  life for people with disabilities and the elderly 
(Khodadadi et al., 2016). IoT systems, such as networked 
vehicles and intelligent traffic systems, are moving 
towards smart cities, reducing congestion and energy 
consumption. However, IoT also presents challenges that 
need to be addressed for potential benefits to be realized 
(Rose et al., 2015).

Class Imbalance in Machine Learning
One of  the significant challenges associated with IoT 
is managing the vast amounts of  data generated by 
these devices, which often leads to class imbalance in 
machine learning applications. In machine learning, 
class imbalance is a widespread problem that impacts 
several industries, such as cybersecurity, finance, and 
healthcare (Dogra et al., 2022). Class imbalance is a major 
difficulty in IoT because data collecting is naturally biased 
towards typical operational conditions (Zhou et al., 2022). 
Imbalanced datasets are those where one of  two possible 
outcomes is rare (Tyagi & Mittal, 2020). A classification 
model’s performance depends on the training dataset’s 
quality and quantity (Hanskunatai, 2018). In imbalanced 
datasets with two-valued classes, accuracy may not clearly 
represent classification results. In applications like disease 
detection and intrusion detection, it is more important 
to correctly predict the minority class (Tyagi & Mittal, 
.2020). The very visible presence of  a class imbalance is 
depicted in Figure 1. 
Class imbalance presents particular difficulties in the 
IoT because of  the type of  data that IoT devices and 
sensor networks produce (Ullah & Mahmoud, 2021). 

Figure 1: Class Imbalance Among Different Datasets 
Source: Author

Class imbalance in IoT information has been studied, 
especially in applications for environmental monitoring, 
predictive maintenance, and anomaly detection (Coelho 
et al., 2022; Fahim & Sillitti, 2019). According to certain 

studies, undersampling strategies might exclude important 
information from the majority class, while oversampling 
could cause overfitting or injecting noise into the data 
(Koziarski et al., .2019; Sáez et al., 2016). To address class 



Pa
ge

 
18

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

imbalance issues in machine learning, the dataset itself  
or the learning methods of  the underlying algorithm can 
be tuned.

Handling Approaches
At the algorithm level, approaches like Ada boosting, 
mapping, and cost-sensitive learning can be used to 
tune the classifier’s results. Data level-based imbalance 
handling involves equating the occurrence of  both classes 
algorithmically to improve the imbalance ratio (Tyagi & 
Mittal, 2020). Several studies have looked into ways to 
address the class disparity, and resampling techniques 
have become prominent approaches. By creating 
synthetic samples, oversampling techniques like Adaptive 
Synthetic Sampling (ADASYN) and Synthetic Minority 
Over-Sampling Technique (SMOTE) seek to boost the 
representation of  minority class instances (Huang, 2015; 
Tarawneh et al., .2020). To rebalance class distributions, 
under-sampling techniques, on the other hand, require 
lowering the quantity of  majority class samples (Abdi 
& Hashemi, .2015). In order to produce a balanced 
dataset, hybrid approaches use both under-sampling and 
oversampling methods.
Several machine learning methods may be used to create 
predictive data models. The model’s accuracy depends on 
how well it can identify the positive class and how well 
it can predict a negative class (Fisher et al., 2019). The 
categorisation rate of  the two classes mentioned above 
has been verified, even if  a model provides 90% accuracy. 
Unbalanced data sets can cause skewed proportions 
between groups, necessitating preprocessing sample 
techniques, algorithmic approaches, or a bot to shift the 
model for sustainable analysis.
ADASYN, CoSen modelling, SMOTE, under- and 
over-sampling, and SMOTE have been commonly 
used solutions. A balanced dataset was produced by 
undersampling, which removes the sample of  the 
dominant class. The loss of  important information was 
ascribed to the dataset’s undersampling. Conversely, over-
sampling attempted to balance the dataset by making 
duplicates of  the pre-existing dataset. It could be arbitrary 

duplicates of  the data subset., leading to overfitting of  the 
model, which is often computationally costly. Instead of  
adding new data samples to the minority class or replacing 
the current samples, the SMOTE-based approach 
artificially produces the sample data. The SMOTE-based 
method faces a problem due to the undesirable addition 
of  noise to the dataset.
The study has focused on adding knowledge on class 
imbalance mitigation in IoT data analysis by comparing 
resampling approaches in IoT datasets. Advancements 
in IoT data analytics have created more durable and 
scalable IoT systems. The study aims to identify 
practical approaches for enhancing the performance and 
dependability of  machine learning models used in IoT 
applications through empirical evaluation and methodical 
comparison.

RESEARCH METHODOLOGY
This research has extensively utilised an experimental 
design to comprehensively contrast resampling techniques 
that were applied to deal with the problem of  class 
imbalance in the presence of  noise data. By leveraging 
resampling methods alongside comparative analysis, the 
dataset was divided into two subsets: a training dataset 
through which the model was trained and a testing dataset 
through which the performance of  the model was tested. 
This experimental setting provided a systematic means 
of  testing the impact of  various resampling schemes on 
classifier performance.

Comparison of  RF and SVC Classifiers
The study includes a comparative investigation of  base 
classifiers, namely the Random Forest Classifier (RF) and 
the Support Vector Classifier (SVC). This comparison 
assesses their effectiveness in resolving class disparity in 
the context of  multi-class issues. According to the study’s 
findings, the RF classifier performs better than other 
basic classifiers in reducing the difficulties caused by class 
imbalance in multi-class situations. The predictive model is 
an RF classifier because of  its solid performance history.

Figure 2: Performance Evaluation of  Base Classifiers on the IoT_Modbus Dataset



Pa
ge

 
19

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

Dataset Selection
In selecting specific datasets for this study, various IoT 
datasets spanning several application areas have been 
explored, selecting those that could demonstrate a 
substantial class imbalance. The main objective of  the 
data selection process was to ensure a comprehensive 
analysis of  specific datasets with various attributes, 
including considering datasets of  various sizes, imbalance 
ratios, and feature space dimensionality.

Preprocessing
Standard data preparation procedures have been 
implemented to ensure the datasets are appropriately 
cleaned, normalised, and subjected to feature 
engineering techniques. This process aims to enhance 
data quality and consistency by effectively addressing 
noise, outliers, and missing values. Resampling strategies 
were employed to rebalance class distributions within 
the IoT datasets, including random under-sampling 
and oversampling techniques such as SMOTE and 
ADASYN. By generating balanced training sets through 
resampling, the models were trained and evaluated more 
effectively, mitigating the impact of  class imbalance and 
improving the overall performance of  the predictive 
models.

Selection of  Base Classifier
Two popular SVC (support vector classifier) and RF 
(random forest classifier) classifiers were selected for 
the study. However, the choice of  specific classifiers has 
depended on the specific properties of  IoT data and its 
function. 

Model Training and Evaluation
The original unbalanced and the resampled datasets 
have been utilised to train the CSV and RF Classifiers. 
Stratified cross-validation has been applied to ensure 
unbiased performance evaluation and mitigate the impact 
of  dataset imbalance during model assessment. Standard 
assessment metrics such as ROC curve (AUC), precision, 
recall, and F1-score have been employed to evaluate 
the performance of  each classifier. Additionally, the 
effectiveness of  SVC and RF classifiers in handling class 
imbalance within the IoT datasets has been compared 
across several resampling methods.

Statistical Analysis
The performance of  SVC and RF classifiers on original 
and resampled datasets has been evaluated by statistical 
methods such as t-tests or Wilcoxon signed-rank tests. 
The outcomes of  the statistical analysis were performed 
to identify significant variations in the performance of  
classifiers. Additionally, the analysis enabled the detection 
of  how resampling methods affect the performance of  
the classifiers.

Sensitivity Analysis 
Sensitivity analysis has been employed to assess the 
resilience of  SVC and RF classifiers to variations in 
dataset properties, encompassing changes in feature space 
dimensionality, dataset size, and class imbalance ratio. By 
systematically varying these properties, the study aimed 
to understand how the classifiers’ performance adapts 
to different data configurations. Additionally, an analysis 
was conducted to investigate the impact of  algorithmic 
decisions and hyperparameter settings on classifier 
performance within diverse resampling scenarios. This 
analysis will provide insights into the robustness of  SVC 
and RF classifiers across various conditions, enabling a 
comprehensive evaluation of  their suitability for handling 
class imbalance and other challenges inherent in IoT 
datasets.

Discussion and Interpretation
The discussion and interpretation part of  the study has 
utilised the analysis of  experimental findings based on 
the comparative performance evaluation of  SVC and RF 
classifiers while employing various resampling strategies. 
The advantages and disadvantages of  each classifier in 
managing the class imbalance have been discussed based 
on the results that have affected IoT data analysis. Key 
factors like interpretability, computational efficiency, and 
model resilience have also been considered to ensure 
effective performance. The study has been summarised 
with possible directions for further studies, such as 
investigating hybrid or ensemble methodologies to enhance 
classifier performance in imbalanced IoT datasets.

Resampling Methods and Model Evaluation Strategy
A collection of  resampling methods, such as “No 
Resampling,” “ROS,” “RUS,” “SMOTE,” and 
“ADASYN,” have been presented to solve class imbalance 
and evaluate its impact on the overall performance of  the 
model. This stage involved determining if  resampling was 
necessary and considering the “No Resampling” option 
to comprehend the impact of  class imbalance on model 
performance. A specific OvO classifier has been utilised 
for multi-class classification, as it was well-suited for the 
situations where several classes were present and can 
be trained efficiently with both original and resampled 
datasets. 
The model’s performance was evaluated during the 
OvO’s training on the chosen dataset, regardless of  
Resampling. After the training phase, predictions were 
created for the test set, and accuracy scores were carefully 
determined. Furthermore, individualised confusion 
matrices were constructed for each resampling scenario, 
comprehensively evaluating the model’s performance 
across various resampling techniques. A visual 
representation of  this methodology has been presented 
in Figure 3.



Pa
ge

 
20

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

RESULTS AND DISCUSSION
Analysis of  Base Classifiers and Resampling 
Strategies in IoT Datasets
A comprehensive analysis of  base classifiers and 
resampling strategies was conducted to identify key 
findings regarding class imbalance in IoT datasets. The 
Random Forest Classifier (RF) and the Support Vector 
Classifier (SVC) performed biasedly, favouring the 
majority class with superior accuracy, precision, and 
recall on the original unbalanced datasets. The study 
revealed a trade-off  between recall and accuracy, with 
SVC demonstrating better recall but poorer precision 
than RF. However, RF regularly outperformed SVC in 
terms of  total F1-score on datasets that were not evenly 
distributed. 
The successful mitigation of  class imbalance resampling 
strategies led to better performance across various 
assessment measures for both classifiers. SMOTE 
and ADASYN oversampling techniques significantly 
improved memory for the minority class, reducing and 
alleviating the imbalance-induced bias in predictions.
Furthermore, sensitivity to the minority class was 
enhanced, but overall accuracy was decreased when 
random undersampling approaches were utilised. 
Although RF consistently outperformed SVC in various 
conditions, demonstrating greater overall accuracy, 
precision, recall, and F1-score in the comparison study of  
SVC and RF resampled datasets. RF also outperformed 
SVC in sensitivity to minority classes and produced a 

stronger recall-to-precision ratio.

Statistical Analysis of  Base Classifier Performance
The statistical analysis of  the study revealed that RF 
outperformed SVC in managing class imbalance to IoT 
datasets, highlighting the significance of  its performance. 
Additionally, various dataset properties such as imbalance 
ratio, size, and feature space dimensionality affected 
the efficacy of  resampling algorithms differently, with 
RF classifiers demonstrating greater resistance to these 
fluctuations than SVC classifiers.
These findings of  the study have emphasised the 
significance of  selecting appropriate resampling techniques 
and classifiers for IoT datasets, thereby enhancing the 
development of  robust and reliable predictive models for 
diverse applications. In this study, the performance of  a 
classification model using various resampling techniques 
was assessed to determine class imbalance in the dataset. 
The classification reports associated with each technique 
yield valuable insights into the model’s precision, recall, 
F1 scores, and accuracy across various classes. These 
results offered a comprehensive perspective on how 
diverse resampling methods influence the model’s ability 
to accurately classify instances across various categories 
(Collell et al., 2018; Qawqzeh et al., 2020). 
The subsequent sections will delve deeper into the 
implications of  these results and their significance in 
selecting the most appropriate resampling technique. 
Figures 4 and 5 represent the classification reports for the 

Figure 3: Illustration of  Resampling Techniques for Class Imbalance Handling in IoT Environment

Figure 4: IoT_Modbus dataset classification reports of  the used resampling techniques



Pa
ge

 
21

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

resampling techniques utilised in this study, focusing on 
both the IoT_Modbus and IoT_GPS_Tracker datasets.
The comprehensive analysis of  the model highlighted the 
impact of  diverse resampling techniques on classification 
performance across varied classes, facilitating a deeper 
understanding of  their implications and avenues for 
potential enhancements. Previous research (Azlim & 
Ahamed, .2023; Jiang et al., 2023; Qawqzeh et al., 2023; 
Rezvani & Wang, 2023) have emphasised the symbiotic 
relationship between resampling techniques and the 
choice of  classification methods, underscoring the 
necessity for synergy to maximise beneficial outcomes.

Absence of  Resampling
In the absence of  Resampling, the model achieved 98% 
notable accuracy. While demonstrating perfect precision 
and high recall for the “Injection” class and a commendable 
F1-score for “Scanning,” lower F1-scores for “XSS” and 
“Backdoor” indicated areas for improvement.

Random Over-Sampling (ROS)
Maintaining a consistent accuracy of  98%, ROS 
exhibited strengths in perfect precision and high recall 
for the “Injection” class, along with a substantial F1-
score for “Scanning.” However, comparatively lower F1 

scores for “Password” and “XSS” suggested potential 
areas for enhancement. ROS effectively addressed class 
distribution imbalance.

Random Under-Sampling (RUS)
Despite balancing class distribution, RUS lead to a 
reduced accuracy of  84%. While achieving a perfect F1-
score for the “Injection” class, the model performance 
was significantly weakened across other classes, resulting 
in low precision and recall.

SMOTE (Synthetic Minority Over-Sampling 
Technique)
SMOTE achieved a commendable accuracy of  99%, 
maintaining excellent precision and recall for “Normal” 
and “Injection” classes, resulting in high F1 scores. 
However, relatively lower F1 scores for “XSS” and 
“Backdoor” suggested potential areas for improvement. 
SMOTE effectively addressed class imbalance by 
generating synthetic instances.

ADASYN (Adaptive Synthetic Sampling)
Similarly achieving a high accuracy of  99%, ADASYN 
sustained robust precision and recall for “Normal” 
and “Injection” classes, resulting in high F1 scores. 

Figure 5: IoT_GPS_Tracker dataset classification reports of  the used resampling techniques

Figure 6: Heat maps showcase the resampling techniques employed on the IoT_Modbus dataset.



Pa
ge

 
22

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

Nevertheless, there was an area for enhancement in F1-
scores for “XSS” and “Backdoor,” signifying potential 
improvements. ADASYN effectively mitigates class 
imbalance through adaptive synthetic sampling. This 
analysis offered insights into the relative effectiveness of  
each resampling approach concerning precision, recall, 
F1 scores, and accuracy across different classes. RUS’s 
overall performance has been compromised despite its 
improvement in the “Injection” class. 
These findings informed conclusions regarding the 
suitability of  resampling techniques in addressing class 
imbalance within a multi-class scenario. Heat maps 
were generated to represent the resampling techniques 
employed on the IoT_Modbus dataset, as shown in 
Figures 6 and 7.

CONCLUSION
The study examined the impact of  resampling techniques 
on classification models in class-imbalanced IoT datasets. 
It found that Support Vector Classifier (SVC) and Random 
Forest Classifier (RF) performed biasedly on unbalanced 
datasets, highlighting the issue of  class imbalance in 
machine learning tasks. Resampling strategies improved 
the performance of  SVC and RF classifiers, with hybrid 
approaches like SMOTE and oversampling techniques 
like ADASYN enabling rebalancing class distributions 
and enhancing model performance. RF consistently 
outperformed SVC in resampling scenarios, achieving 
superior accuracy, precision, recall, and F1-score. 
The study emphasized the importance of  resampling 
techniques in scenarios marked by class imbalance to 
enhance the accuracy and reliability of  classification 
models in practical applications. Future research could 
explore hyperparameter tuning’s effects on model 
performance and explore the applicability of  these 
techniques in domains like cybersecurity, fraud detection, 
and medical diagnosis.

Future Implications
This subsequent study may investigate several directions 

to expand our comprehension and improve the 
usefulness of  resolving class imbalance in IoT datasets. 
Exploring hybrid or adaptive resampling methodologies 
might mitigate trade-offs observed in current techniques 
and potentially boost overall classification performance. 
Examine how well ensemble learning methods-like 
bagging, boosting, or stacking-work with resampling 
techniques to enhance model robustness and performance 
in unbalanced IoT datasets. By combining the advantages 
of  several classifiers and resampling strategies, ensemble 
approaches may improve generalisation and prediction 
accuracy. 
The study explores adaptive resampling techniques that 
can dynamically adjust to changes in the data distribution. 
Ensuring the ongoing efficacy of  class imbalance 
mitigation strategies in practical applications may 
require creating algorithms that can recognise and react 
to idea drift, data drift, or changing class distributions 
in IoT datasets. The study suggests that exploring the 
compatibility between advanced classification algorithms 
and resampling methods is a promising direction. This 
exploration could unveil enhanced performance in 
complex multi-class scenarios, presenting opportunities 
for more robust and accurate models.
Moreover, examining the practical implications and 
robustness of  these techniques in real-world scenarios, 
particularly in domains where accurate classification is 
imperative, would be instrumental. This includes rigorous 
testing and validation of  these techniques in operational 
settings to gauge their effectiveness and feasibility beyond 
controlled experimental setups. Validate the effectiveness 
of  the identified resampling techniques and classifiers 
through deployment in real-world IoT environments. 
Conduct extensive evaluation and monitoring of  model 
performance under practical conditions, considering 
scalability, reliability, and interpretability factors. 
Case studies and field trials in diverse IoT domains 
could provide valuable insights into the applicability 
and impact of  class imbalance mitigation strategies in 
real-world settings. The project proposes frameworks 

Figure 7: Heat maps showcase the resampling techniques employed on the IoT_GPS_Tracker dataset



Pa
ge

 
23

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

and automated methods for selecting suitable classifiers, 
hyperparameters, and resampling techniques for IoT 
dataset’s properties. The model creation process might 
be streamlined by automated model selection and 
hyperparameter tweaking, allowing practitioners to 
quickly find and implement efficient predictive models in 
Internet of  Things applications.
The findings of  this study provide a foundation for future 
research, emphasising the need for tailored techniques and 
their practical applications in addressing class imbalance 
within the dynamic landscape of  IoT datasets. The 
enhancement of  the state-of-the-art in-class imbalance 
mitigation strategies for IoT datasets by addressing 
these future research objectives will eventually improve 
the predictive modelling’s performance, scalability, and 
reliability in various IoT applications.

Acknowledgment
We would like to express our gratitude to Yousef  
Qawqzeha from the University of  Fujairah for his 
valuable contributions to this research. His expertise 
and dedication greatly enriched the development of  this 
study. We also acknowledge the support provided by 
his Corresponding Author Email (YousefQawqzehaa@
outlook.com) and his ORCID (0000-0001-7774-062x) 
throughout the research process.

References
Abdi, L., & Hashemi, S. (2015). To combat multi-class 

imbalanced problems by means of  over-sampling 
techniques. IEEE transactions on Knowledge and Data 
Engineering, 28(1), 238-251. 

Atuhurra, J., Hara, T., Zhang, Y., Sasabe, M., & Kasahara, 
S. (2024). Dealing with Imbalanced Classes in Bot-
IoT Dataset. arXiv preprint arXiv:2403.18989. 

Azlim Khan, A. K., & Ahamed Hassain Malim, N. 
H. (2023). Comparative Studies on Resampling 
Techniques in Machine Learning and Deep Learning 
Models for Drug-Target Interaction Prediction. 
Molecules, 28(4), 1663. 

Coelho, D., Costa, D., Rocha, E. M., Almeida, D., & Santos, 
J. P. (2022). Predictive maintenance on sensorized 
stamping presses by time series segmentation, 
anomaly detection, and classification algorithms. 
Procedia Computer Science, 200, 1184-1193. 

Collell, G., Prelec, D., & Patil, K. R. (2018). A simple 
plug-in bagging ensemble based on threshold-moving 
for classifying binary and multiclass imbalanced data. 
Neurocomputing, 275, 330-340. 

Dogra, V., Verma, S., Verma, K., Jhanjhi, N. Z., Ghosh, 
U., & Le, D.-N. (2022). A comparative analysis of  
machine learning models for banking news extraction 
by multiclass classification with imbalanced datasets 
of  financial news: challenges and solutions.

Fahim, M., & Sillitti, A. (2019). Anomaly detection, 
analysis and prediction techniques in iot environment: 
A systematic literature review. IEEE Access, 7, 81664-
81681. 

Fisher, A., Rudin, C., & Dominici, F. (2019). All models 
are wrong, but many are useful: Learning a variable’s 
importance by studying an entire class of  prediction 
models simultaneously. Journal of  Machine Learning 
Research, 20(177), 1-81. 

Hanskunatai, A. (2018). A new hybrid sampling approach 
for classification of  imbalanced datasets. 2018 3rd 
International Conference on Computer and Communication 
Systems (ICCCS), 

Huang, P. J. (2015). Classification of  imbalanced data 
using synthetic over-sampling techniques. University 
of  California, Los Angeles. 

Jiang, X., Wang, J., Meng, Q., Saada, M., & Cai, H. (2023). 
An adaptive multi-class imbalanced classification 
framework based on ensemble methods and deep 
network. Neural Computing and Applications, 35(15), 
11141-11159. 

Johnson, J. M., & Khoshgoftaar, T. M. (2019). Survey on 
deep learning with class imbalance. Journal of  Big Data, 
6(1), 1-54. 

Khodadadi, F., Dastjerdi, A. V., & Buyya, R. (2016). 
Internet of  things: an overview. Internet of  things, 3-27. 

Koziarski, M., Krawczyk, B., & Woźniak, M. (2019). 
Radial-based oversampling for noisy imbalanced data 
classification. Neurocomputing, 343, 19-33. 

Koziarski, M., Woźniak, M., & Krawczyk, B. (2020). 
Combined cleaning and resampling algorithm for 
multi-class imbalanced data with label noise. Knowledge-
Based Systems, 204, 106223. 

Kumar, A., Sharma, S., Singh, A., Alwadain, A., Choi, B.-
J., Manual-Brenosa, J., Ortega-Mansilla, A., & Goyal, 
N. (2021). Revolutionary strategies analysis and 
proposed system for future infrastructure in internet 
of  things. Sustainability, 14(1), 71. 

Nixon, C., Sedky, M., & Hassan, M. (2019). Practical 
application of  machine learning based online intrusion 
detection to internet of  things networks. 2019 IEEE 
Global Conference on Internet of  Things (GCIoT). 

Nord, J. H., Koohang, A., & Paliszkiewicz, J. (2019). 
The Internet of  Things: Review and theoretical 
framework. Expert Systems with Applications, 133, 97-
108. 

Obaid, W., & Nassif, A. B. (2022). The effects of  
resampling on classifying imbalanced datasets. 2022 
Advances in Science and Engineering Technology International 
Conferences (ASET).

Paisitkriangkrai, S., Shen, C., & van den Hengel, A. 
(2013). A scalable stagewise approach to large-margin 
multiclass loss-based boosting. IEEE transactions on 
neural networks and learning systems, 25(5), 1002-1013. 

Pal, D., Funilkul, S., Charoenkitkarn, N., & Kanthamanon, 
P. (2018). Internet-of-things and smart homes for 
elderly healthcare: An end user perspective. IEEE 
Access, 6, 10483-10496. 

Peng, H., Wu, C., & Xiao, Y. (2023). CBF-IDS: Addressing 
Class Imbalance Using CNN-BiLSTM with Focal 
Loss in Network Intrusion Detection System. Applied 
Sciences, 13(21), 11629. 



Pa
ge

 
24

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 16-24, 2025

Powroźnik, P., Szcześniak, P., & Piotrowski, K. (2021). 
Elastic energy management algorithm using IoT 
technology for devices with smart appliance 
functionality for applications in smart-grid. Energies, 
15(1), 109. 

Pramanik, P. K. D., Upadhyaya, B. K., Pal, S., & Pal, 
T. (2019). Internet of  things, smart sensors, and 
pervasive systems: Enabling connected and pervasive 
healthcare. In Healthcare data analytics and management 
(pp. 1-58). Elsevier. 

Qawqzeh, Y. K., Alourani, A., & Ghwanmeh, S. (2023). 
An improved breast cancer classification method 
using an enhanced AdaBoost classifier. International 
Journal of  Advanced Computer Science and Applications, 
14(1). 

Qawqzeh, Y. K., & Ashraf, M. (2023). A Fraud Detection 
System Using Decision Trees Classification in 
An Online Transactions. Proceedings of  the 2023 
12th International Conference on Software and Computer 
Applications.

Qawqzeh, Y. K., Bajahzar, A. S., Jemmali, M., Otoom, M. 
M., & Thaljaoui, A. (2020). Classification of  diabetes 
using photoplethysmogram (PPG) waveform 
analysis: logistic regression modeling. BioMed Research 
International, 2020.

Rezvani, S., & Wang, X. (2023). A broad review on class 
imbalance learning techniques. Applied Soft Computing, 
110415. https://doi.org/10.1016/j.asoc.2023.110415

Rose, K., Eldridge, S., & Chapin, L. (2015). The internet 
of  things: An overview. The internet society (ISOC), 
80(15), 1-53. 

Sáez, J. A., Krawczyk, B., & Woźniak, M. (2016). Analyzing 
the oversampling of  different classes and types of  
examples in multi-class imbalanced datasets. Pattern 
Recognition, 57, 164-178. 

Tanha, J., Abdi, Y., Samadi, N., Razzaghi, N., & 
Asadpour, M. (2020). Boosting methods for multi-
class imbalanced data classification: an experimental 
review. Journal of  Big Data, 7, 1-47. 

Tarawneh, A. S., Hassanat, A. B., Almohammadi, K., 
Chetverikov, D., & Bellinger, C. (2020). Smotefuna: 

Synthetic minority over-sampling technique based on 
furthest neighbour algorithm. IEEE Access, 8, 59069-
59082. 

Tyagi, S., & Mittal, S. (2020). Sampling approaches for 
imbalanced data classification problem in machine 
learning. Proceedings of  ICRIC 2019: Recent 
innovations in computing, 

Ullah, I., & Mahmoud, Q. H. (2021). A framework for 
anomaly detection in IoT networks using conditional 
generative adversarial networks. IEEE Access, 9, 
165907-165931. 

Varotto, G., Susi, G., Tassi, L., Gozzo, F., Franceschetti, 
S., & Panzica, F. (2021). Comparison of  resampling 
techniques for imbalanced datasets in machine 
learning: application to epileptogenic zone localization 
from interictal intracranial EEG recordings in patients 
with focal epilepsy. Frontiers in Neuroinformatics, 15, 
715421. 

Wanasinghe, T. R., Gosine, R. G., James, L. A., Mann, G. 
K., De Silva, O., & Warrian, P. J. (2020). The internet 
of  things in the oil and gas industry: a systematic 
review. IEEE Internet of  Things Journal, 7(9), 8654-
8673. 

Wang, S., & Yao, X. (2012). Multiclass imbalance problems: 
Analysis and potential solutions. IEEE Transactions on 
Systems, Man, and Cybernetics, Part B (Cybernetics), 42(4), 
1119-1130. 

Welvaars, K., Oosterhoff, J. H., van den Bekerom, M. P., 
Doornberg, J. N., van Haarst, E. P., OLVG Urology 
Consortium, & R, t. M. L. C. v. d. Z. J. v. A. G. L. 
B. H. M. K. P. B. L. v. d. K. A. M. W. P. (2023). 
Implications of  resampling data to address the class 
imbalance problem (IRCIP): an evaluation of  impact 
on performance between classification algorithms in 
medical data. JAMIA open, 6(2), ooad033. 

Zhou, X., Hu, Y., Wu, J., Liang, W., Ma, J., & Jin, Q. (2022). 
Distribution bias aware collaborative generative 
adversarial network for imbalanced deep learning 
in industrial IoT. IEEE Transactions on Industrial 
Informatics, 19(1), 570-580. 


