EUROPEAN JOURNAL OF PURE AND APPLIED MATHEMATICS
2025, Vol. 18, Issue 2, Article Number 6087
ISSN 1307-5543 – ejpam.com
Published by New York Business Global

Towards Accurate Fake News Detection: Evaluating
Machine Learning Approaches and Feature Selection

Strategies

Mutaz A. B. Al-Tarawneh1,∗, Ashraf Al-Khresheh2,, Omar Al-irr1,, Ajla Kulaglic1,,
Kassem Danach3,, Hassan Kanj1,, Ghayth AlMahadin4

1 College of Engineering and Technology, American University of the Middle East, Egaila
54200, Kuwait
2 Computer Science Department, Tafila Technical University,Tafila, Jordan
3 Basic and Applied Sciences Research Center, Al Maaref University, Beirut, Lebanon
4 Data Science Department, Mutah University, Karak, 61710, Jordan

Abstract. The rapid spread of fake news in the digital age poses significant challenges, neces-
sitating effective detection methods. This study presents a comprehensive evaluation of various
ensemble and machine learning classifiers, combined with different feature selection techniques, to
improve the accuracy and reliability of the detection of fake news. Using the TruthSeeker dataset,
this research examines feature selection methods such as Recursive Feature Elimination (RFE), Se-
lectKBest, Principal Component Analysis (PCA), and Genetic Algorithms (GA), analyzing their
impact on model performance. Key metrics such as accuracy, precision, recall, F1 score, and
AUC-ROC were used to assess the effectiveness of each classifier. The results reveal that ensemble
methods, particularly Random Forest (RF) and Gradient Boosting, demonstrate superior perfor-
mance, achieving high accuracy and AUC-ROC scores. Moreover, feature selection techniques like
RFE and SelectKBest significantly improve model outcomes by optimizing the feature set, while
PCA is less effective in this context. This study highlights the importance of integrating robust
classifiers with optimal feature selection methods to improve the efficacy of fake news detection
systems.

2020 Mathematics Subject Classifications: 68T01

Key Words and Phrases: Fake news detection, Machine learning, Feature selection, Truthseeker

∗Corresponding author.
DOI: https://doi.org/10.29020/nybg.ejpam.v18i2.6087

Email addresses: mutaz.al-tarawneh@aum.edu.kw (M. Al-Tarawneh),
khashraf@ttu.edu.jo (A. Al-Khresheh), omar.alirr@aum.edu.kw (O. Alirr),
ajla.kulaglic@aum.edu.kw (A. Kulaglic), kassem.danach@iul.edu.lb (K. Danach),
hassan.kanj@aum.edu.kw (H. Kanj), ghayth.mahadin@mutah.edu.jo (G. AlMahadin)

https://www.ejpam.com 1 Copyright: © 2025 The Author(s). (CC BY-NC 4.0)



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 2 of 39

1. Introduction

The advent of the digital age has transformed the way information is accessed and dis-
seminated, leading to unprecedented opportunities for knowledge sharing. However, this
increased accessibility has also facilitated the widespread circulation of misinformation
and fake news. Fake news, characterized as deliberately false or misleading content pre-
sented in the guise of authentic news, has emerged as a critical challenge with far-reaching
consequences. It influences public perception, shapes political decisions, and can even lead
to social instability. The difficulty of distinguishing genuine news from fabricated stories
is exacerbated by the rapid and extensive distribution of information through social media
and digital platforms [1].

The repercussions of fake news extend into the physical world, eroding trust in tra-
ditional media and impacting real-world events. For example, the 2016 US presidential
election saw a proliferation of misleading stories that may have affected voter behavior
and the final outcome [2]. Similarly, during the 2019 Indian general election, targeted
misinformation campaigns spread false narratives and incited division among voters [3].
The COVID-19 pandemic further highlighted the dangers of false news, with widespread
misinformation about the virus, treatments, and vaccines creating significant public health
risks [4]. These instances underscore the pressing need for robust strategies to identify
and mitigate fake news.

Machine learning (ML) has become a cornerstone in addressing this challenge, offering
advanced methods to identify and classify fake news. By analyzing extensive datasets
with sophisticated algorithms, ML models can detect patterns and attributes indicative
of misinformation [5]. The fake news detection process typically involves several stages:
data collection, feature extraction, model training, and evaluation. Among these, feature
selection plays a pivotal role in pinpointing the most relevant attributes in the dataset,
thereby enhancing model accuracy and efficiency.

Feature selection is crucial to improve the performance of the ML model, as it reduces
complexity and increases interpretability [6]. Several techniques have been developed for
this purpose. Recursive Feature Elimination (RFE) incrementally removes less relevant
features based on model weights to refine the set of features [7]. SelectKBest applies sta-
tistical tests to identify the key features that most closely correlated with the target vari-
able [8]. Principal Component Analysis (PCA) transforms existing features into a smaller
set of components, retaining maximum variance [9]. Furthermore, Genetic Algorithms
(GAs) offer a global optimization approach by mimicking natural selection processes to
identify the optimal set of characteristics in complex search spaces [10].

This research presents a comprehensive evaluation of machine learning methods for
detecting fake news. Explores the performance of several individual classifiers, includ-
ing Decision Trees (DT), Näıve Bayes (NB), Support Vector Machine (SVM), K-Nearest
Neighbors (KNN), Random Forest (RF), and Multilayer Perceptron (MLP), along with
ensemble approaches like Bagging, Boosting, and Stacking. An emphasis is placed on the
influence of various feature selection methods on model effectiveness. A robust evalua-
tion framework that employs metrics such as accuracy, precision, recall, F1 score, and



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 3 of 39

AUC-ROC ensures a thorough analysis of the model capabilities.
A key asset in this research is the TruthSeeker dataset, which encompasses a decade

worth of news articles, tweets, and social media content. This diverse dataset supports
a detailed analysis by providing textual, lexical, and metadata features. Inclusion of
metadata, such as user engagement statistics, allows the detection of subtle patterns that
textual analysis alone may miss. The temporal and content diversity of the dataset ensures
that the models are evaluated across a wide range of real-world scenarios.

By systematically comparing individual classifiers and ensemble methods alongside
various feature selection techniques, this study sheds light on the optimal strategies for
effective fake news detection. The contributions of this work include the following.

• Comprehensive Evaluation: A detailed assessment of multiple machine learning
models for the detection of fake news using various performance metrics such as
accuracy, precision, recall, F1 score and AUC-ROC.

• Impactful Feature Selection: Analysis of the effects of feature selection methods,
including RFE, SelectKBest, PCA, and GA, on model performance.

• Rich Dataset Utilization: Utilization of the TruthSeeker dataset, offering diverse
and temporally extensive data to ensure robust model evaluation.

• Holistic Analysis: Examination of textual, lexical, and metadata features to cap-
ture nuanced patterns in the detection of fake news.

• Comparative Analysis: Insights into the performance of individual classifiers
versus ensemble methods to identify optimal detection strategies.

The remainder of this paper is structured as follows. Section 2 reviews related work,
Section 3 details the methodology, including data collection, machine learning techniques,
feature selection, and evaluation. Section 4 discusses the results and findings, and Section
5 concludes the study with key findings.

2. Literature Review

Fake news detection has emerged as a critical research area due to the pervasive spread
of misinformation across digital platforms. In recent years, several studies have explored
ML techniques to tackle this pressing issue. This review of the literature discusses no-
table work in the field, highlighting their methodologies and comparing them with current
research.

The study in [11] presented BiL-FaND, an ensemble-based system for detecting fake
news in English and Urdu. Combining multilingual BERT, long-short-term memory
(LSTM) models, and a ResNet-101 with GRU for multimedia analysis, the system achieved
an overall notable accuracy of 92.07%. The approach demonstrated strong performance
in linguistic analysis, pattern recognition, and multimedia content evaluation. However,
the study does not explore feature selection methods.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 4 of 39

The authors of [12] developed a robust fake news detection approach using WELFake,
FakeNewsNet, and FakeNewsPrediction datasets. FastText embeddings were combined
with ML and deep learning (DL) methods, including a hybrid CNN-LSTM model that
achieved high performance (accuracy and F1 scores of 99%, and 97%). Transformer-based
models (BERT, XLNet, RoBERTa) were also optimized for superior syntactic manage-
ment. Explainable AI techniques (LIME, LDA) provided insight into model decisions.
However, the study does not explore feature selection methods.

The work in [13] addressed the spread of COVID-19 misinformation by developing a
Convolutional Neural Network (CNN)-based deep learning model using word embeddings.
The optimal CNN architecture was determined through grid search. The effectiveness
of the model was evaluated against various state-of-the-art ML algorithms, with CNN
achieving the highest performance: 96.19% mean accuracy, 95% mean F1 score, and 0.985
AUC. However, they did not explore feature selection methods.

The study in [14] examined the impact of fake news on society and presented a com-
parative analysis of ML methods to detect fake news. The models evaluated included NB,
DT classifier, RF, and logistic regression (LR). The results indicated that the DT classifier
achieved the highest accuracy at 99.56%, followed by RF (99. 35%), LR (98. 91%) and
NB (94. 89%). Recall results were also better for DT and RF compared to LR and NB. In
particular, the study focused on model comparison rather than exploring feature selection
methods.

The research proposed in [15] introduced an innovative approach to detecting fake
news by analyzing semantic discrepancies between the titles and content of the articles.
Using feature selection, the contents of the articles are summarized, and the vector dis-
tances (Cosine similarity, Jaccard distance, Euclidean distance) between the titles and the
contents are calculated. These distances are combined as features to train the ML and
DL models, achieving almost 99.9% accuracy in three data sets. This method demon-
strates the effectiveness of exploiting semantic differences and underscores the importance
of feature selection in enhancing detection performance.

The approach of [16] addressed the rapid spread of fake news by investigating style-
based detection using text analysis with natural language processing techniques. Various
models were built using word representations (TF-IDF, word2Vec) and ML (e.g., KNN,
NB, LR) and DL (e.g., LSTM) methods on the ISOT dataset. Performance was evaluated
using accuracy, precision, recall, and the F1 score, with the LSTM model achieving the
highest accuracy at 99.2%. Although the study focused on improving word representations
and classification methods, it did not investigate feature selection techniques, which are
also important to optimize model performance.

The study in [17] explored various ML algorithms to identify fake news, with the aim
of determining which algorithm achieved the highest precision. The study found that the
DT algorithm performed the best, achieving an accuracy of 99.97% compared to other
methods. However, the research did not investigate feature selection techniques, which
are also crucial to enhancing model performance.

The study in [18] explored various ML models to classify the veracity of information and
detect fake news on social networks. The analysis used a dataset of approximately 40,000



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 5 of 39

items (20,000 fake and 20,000 real news) and used a variety of models, including SVM,
LR, CatBoost, XGBoost, multinomial NB, and RF. Performance was evaluated using
metrics such as accuracy, precision, recall, F1 score, and more. The deep Auto ViML
model achieved the highest accuracy, precision, recall and F1 score at 99%, while the
hybrid learning model had the best false rejection rate at 71%. The SVM was noted for
its computational efficiency, taking only 0.245 ms for computation. However, the study
did not investigate feature selection techniques, which are also crucial to enhancing model
performance.

The authors of [19] proposed a method that uses ML classification to distinguish be-
tween genuine and fake content. The approach used natural language processing tech-
niques, including the TF-IDF and Word2Vec features, which were optimized using the
RFE algorithm for feature selection. Training and testing were performed on the Kaggle
’Fake News Detection’ dataset, and the results showed that the boosting ensemble meth-
ods outperformed other techniques. The study highlighted the effectiveness of feature
selection through RFE, but did not explore other feature selection methods.

The study in [20] aimed to build a model to detect fake news by analyzing the char-
acteristics of the text. TF-IDF technology was used to convert words into features and
identify the highest-ranking features to distinguish between real and fake news. The study
then adapted ML techniques including LR, DT, gradient boosting and RF, with DT and
Gradient Boosting achieving the highest precision at 99.4% and 99.49%, respectively. Al-
though the study focused on feature extraction through TF-IDF, it did not explore other
feature selection methods, which could have further enhanced model performance.

The research presented in [21] evaluated the effectiveness of three ML algorithms-
Multinomial NB, Passive Aggressive Classifier, and LR-to distinguish between fake and
genuine news articles. Using a balanced dataset, the study processed and vectorized the
text with TF-IDF and assessed algorithms based on precision, recall and F1 score. The
Passive Aggressive Classifier achieved the highest performance with a precision and recall
of 0.99, while LR had an accuracy of 0.98 and Multinomial NB exhibited a robust recall
of 100% but lower precision at 91%, resulting in an accuracy of 95%. The study did not
explore advanced feature selection methods beyond TF-IDF, which could have further
improved the models’ performance.

The paper in [22] addressed the challenge of fake news detection by proposing the
use of various AI and ML methods, combining both individual and ensemble approaches
for automated news classification. The study explored different textual properties to
distinguish fake news from real news and applied multiple ML algorithms to a real-world
dataset for this classification task. The performance of these models was evaluated on
the basis of their accuracy in identifying fake news. However, the study did not discuss
advanced feature selection techniques beyond basic textual properties.

The research in [23] addressed the challenge of fake news detection in Pakistan by
creating a comprehensive dataset from multiple fact-checked news APIs and evaluating it
using a range of ML and DL techniques. The study applied five ML algorithms: NB, KNN,
LR, SVM, and DT, as well as two DL models: CNN and LSTM using GloVe and BERT
embeddings for text representation. The models were assessed based on precision, F1



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 6 of 39

score, accuracy, and recall, with the LSTM model using GloVe embeddings achieving the
highest performance. However, the study did not investigate feature selection methods
but focused on different embeddings and algorithms for classification. In addition, the
research included an analysis of misclassified samples by comparing them with human
judgments.

The study in [24] addressed the challenge of fake news detection by developing a model
that incorporates user characteristics, news content, and social network features based on
social capital. The XGBoost model was used to estimate the importance of the features
and identify key variables for the detection of fake news. Several ML algorithms, including
SVM, RF, LR, Classification and Regression Trees (CART) and Neural Networks (NN),
were applied to classify news articles. A cross-validation step was performed to generalize
the models and prevent overfitting or underfitting. The RF model achieved the highest
prediction rate of approximately 94%, while the NN model had the lowest performance of
about 92.1%. The study did not explore traditional feature selection methods but focused
on derived features based on social capital and estimated their importance to detect fake
news. The work in [25] provided valuable information on the impact of different word
embedding techniques on the performance of machine learning and deep learning models
for the detection of fake news. Although that study focused on embedding-based feature
representations and compared individual classifiers and CNN-based architectures, it did
not consider ensemble learning techniques or feature selection strategies beyond embed-
dings. In contrast, the current study adopts a different research direction by conducting
a comprehensive evaluation of traditional and ensemble classifiers in combination with
multiple feature selection techniques, including Recursive Feature Elimination (RFE), Se-
lectKBest, Principal Component Analysis (PCA), and Genetic Algorithms (GA). This
approach enables a deeper investigation into the relevance and interpretability of features
and offers comparative information on how various selection methods influence classifica-
tion performance, areas not covered in the previous work.

In contrast to these studies, this research presents several distinctive strengths. Firstly,
this study employs a diverse set of individual classifiers, including DT, NB, SVM, KNN,
RF, and MLP, along with ensemble methods such as Bagging, Boosting, and Stacking.
This comprehensive approach ensures a robust evaluation of various algorithms for the
detection of fake news. In addition, the research systematically investigates the impact of
various feature selection methods, including RFE, SelectKBest, PCA, and GA, on model
performance. This detailed examination of feature selection techniques provides valuable
insight into how different methods influence the effectiveness of fake news detection models.
Furthermore, the study uses the TruthSeeker dataset, which offers a rich and varied source
of data comprising textual, lexical, and metadata features. Spanning more than a decade,
this extensive dataset ensures that the models are tested on a wide range of real-world
scenarios. The inclusion of metadata features, such as user behavior and engagement
metrics, allows for a more nuanced analysis of fake news, capturing patterns that may
not be evident from textual analysis alone. This holistic approach improves the accuracy
and robustness of fake news detection models, which separates this study from others that
focus primarily on advanced model architectures or specific feature sets. In summary,



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 7 of 39

while recent studies have made significant contributions to the field of fake news detection
through various ML techniques and datasets, current research distinguishes itself through
a comprehensive evaluation of multiple classifiers and feature selection methods using the
diverse and extensive TruthSeeker dataset. This multifaceted approach provides a more
detailed understanding of effective strategies for detecting fake news and advancing the
state of the art in this critical area of research.

3. Methodology

This section highlights the stages followed in this investigation, including data col-
lection, implementation of machine learning for the detection of fake news, selection of
features and performance evaluation.

3.1. Data Collection

This study employs the TruthSeeker dataset, one of the largest ground-truth collec-
tions for fake news detection on Twitter, which spans from 2009 to 2022 and includes
more than 180,000 labeled tweets. The dataset creation followed a structured multi-phase
annotation process designed to ensure high reliability. Initial real and fake statements
were collected from PolitiFact and keyword-based crawling was used to retrieve related
tweets. The annotation process involved Amazon Mechanical Turk (MTurk), where each
tweet-statement pair was evaluated by three independent Master Turkers. These highly
qualified annotators completed semantic similarity tasks, judging how much a tweet agreed
or disagreed with the source statement using predefined five-class and three-class labeling
schemes. Only responses with a two-thirds majority agreement were retained to ensure
label consistency.This robust process, grounded in active learning and human verification,
makes TruthSeeker a reliable resource for both binary and multiclass fake news detection.

In addition to labels, the TruthSeeker dataset offers enriched information to enhance
its analytical capabilities. Each tweet is supplemented with scores that assess bot likeli-
hood, user credibility, and influence, providing a comprehensive view of the source and its
potential impact. These scores are critical to understanding the mechanisms of automated
misinformation, assessing content credibility, and studying the role of influential users in
spreading false information. Furthermore, the dataset includes rich metadata, such as
user profiles, tweet timestamps, and engagement metrics (likes, retweets, replies), offering
invaluable insights into the spread and impact of fake news across networks.

The TruthSeeker dataset is made available through the Canadian Institute for Cyber-
security (CIC) as an open resource for researchers worldwide. Its high validation standards
and extensive metadata make it indispensable for advancing fake news detection method-
ologies. Researchers can use it to develop and test new machine learning models, explore
patterns in misinformation dissemination, and design interventions to mitigate its spread.
In general, the TruthSeeker dataset is a meticulously validated and richly annotated re-
source that provides a solid foundation to tackle the complex issue of misinformation on
social networks. Tables 1, 2, and 3 summarize the text, lexical, and metadata features of



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 8 of 39

Table 1: Text Features for Fake News Detection

Name Description

Unique Count Total count of unique, complex terms in text

Total Count Overall word count in the text

ORG Percent Percentage of text referencing organizations
NORP Percent Percentage of text mentioning groups or affiliations

GPE Percent Percentage of text related to geopolitical entities

PERSON Percent Percentage of text including references to individuals
MONEY Percent Percentage of text discussing monetary amounts

DATA Percent Percentage of text containing references to dates

CARDINAL Percent Percentage of text using cardinal numbers
PERCENT Percent Percentage of text involving percentages

ORDINAL Percent Percentage of text using ordinal numbers

LAW Percent Percentage of text mentioning legal documents
PRODUCT Percent Percentage of text referring to products

EVENT Percent Percentage of text mentioning specific events
TIME Percent Percentage of text referring to time-related details

LOC Percent Percentage of text with references to locations

WORK OF ART Percent Percentage of text mentioning artistic works
QUANTITY Percent Percentage of text involving quantities

LANGUAGE Percent Percentage of text referring to languages

Max Word Length of the longest term in the text
Min Word Length of the shortest term in the text

Avg Word Length Average length of terms within the text

the dataset, respectively.

3.2. Feature Preprocessing: Standardization

An essential phase in the machine learning workflow is feature preprocessing, with
standardization being one of the most significant techniques applied [26]. Standardization
ensures that all features are rescaled to have a mean of 0 and a standard deviation of
1, which is vital for the effective functioning of many machine learning algorithms. This
transformation is performed independently for each feature using the following formula:

x′ =
x− µ

σ
(1)

Here, x represents the original feature value, µ is the mean of the feature, and σ
denotes the standard deviation of the feature. After applying this formula, each feature is
standardized to have a mean of 0 and a standard deviation of 1.

Standardization is particularly critical for algorithms that are sensitive to the scale of
input data. Models such as Logistic Regression (LR), Support Vector Machines (SVM),
and Neural Networks (NN) rely on consistent feature scaling to ensure effective learn-
ing. Without standardization, algorithms may disproportionately emphasize features with
larger ranges, resulting in suboptimal model performance.

By normalizing the feature scales, these algorithms can focus on identifying the true
relationships between the input features and the target variable, rather than being misled
by differences in feature magnitudes. In addition, standardization improves the numerical
stability and efficiency of optimization algorithms. Standardized features minimize the



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 9 of 39

Table 2: Lexical Features for Fake News Detection

Name Description

Present Verb Count of present tense verbs

Past Verb Count of past tense verbs

Adjectives Total number of adjectives in the text
Pronouns Total number of pronouns in the text

TO’s Frequency of the word ”to”

Determiners Count of determiner words
Conjunctions Total count of conjunctions

Dots Instances of period (.) usage

Exclamations Instances of exclamation marks (!) usage
Question Instances of question marks (?) usage

Ampersand Frequency of ampersand (&) usage

Capitals Count of capitalized letters
Quotes Total instances of quotation marks

Digits Frequency of numerical digits (0-9)
Long Word Freq Frequency of long words

Short Word Freq Frequency of short words

Table 3: Meta-Data Features for Twitter Users

Name Description

Followers Count Total number of followers of the user
Friends Count Total number of friends connected to the user

Favourites Count Total likes given across all user tweets

Statuses Count Overall count of tweets by the user
Listed Count Number of times the user’s tweets are listed

Mentions Total mentions of the user in tweets

Quotes Count of times the user’s tweets were quoted
Replies Count of replies received by the user

Retweets Total retweets of the user’s content

Favourites Count of times the user’s tweets were favorited
Hashtags Number of hashtags used in the user’s tweets

URLs Whether the user’s profile contains a URL
BotScoreBinary Binary indicator of the user being a bot

Cred Score measuring the user’s credibility

Normalized Influence Normalized score of the user’s influence

risk of numerical errors, such as overflow or underflow, that can arise during complex
calculations.

Another significant benefit of standardization is its role in accelerating optimization
algorithms, such as gradient descent. When features are on a uniform scale, the optimiza-
tion process becomes more stable and achieves convergence faster, helping the model to
find the optimal solution efficiently.

3.3. Classification Techniques

This section outlines the classification techniques employed in the study, focusing on
both individual classifiers and ensemble methods. The evaluation process aims to as-
sess their effectiveness and identify the most suitable models for integration into ensemble
frameworks. The classification approaches are categorized into two groups: base classifiers
and ensemble methods, with the latter further divided into homogeneous and heteroge-



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 10 of 39

neous strategies.

3.3.1. Base Classifiers

The study begins by evaluating various individual classifiers to determine their effective-
ness for inclusion in the ensemble. Identifying high-performing models is essential for
constructing robust ensemble methods [27–29].

• Decision Trees (DT) use a tree-like structure to model decision-making processes,
where nodes represent feature tests, branches denote outcomes, and leaves indicate
class labels. The tree is built by partitioning the data based on attributes that
maximize information gain. DTs are easy to interpret and visualize, accommodating
both numerical and categorical features. However, they are susceptible to overfitting,
especially with deep trees or small datasets [30, 31].

• Naive Bayes (NB) is a probabilistic classifier based on Bayes’ theorem, often applied
to text classification tasks. It assumes feature independence given the class, simplifies
computations, and enables efficient training. Despite its simplicity, NB performs well
in areas such as spam detection and sentiment analysis, but struggles with highly
correlated features or non-Gaussian data distributions [32].

• Support Vector Machines (SVMs) are versatile algorithms that separate classes by
finding the hyperplane that maximizes the margin between them. SVMs can handle
linear and non-linear problems using kernel functions such as the linear, polynomial,
and radial basis function (RBF). The regularization parameter C controls the trade-
off between the margin size and the classification errors [33, 34].

• K-Nearest Neighbors (KNN) is a straightforward algorithm that assigns class labels
based on the majority class among the K nearest neighbors. Distance metrics such as
Euclidean or Manhattan are used to find neighbors. KNN is effective for multiclass
classification, but can be computationally expensive for large datasets and is sensitive
to values of K and the dimensionality of features [31, 35].

• Random Forest (RF) builds multiple decision trees on subsets of random data and
aggregates their predictions. This ensemble approach reduces overfitting and vari-
ance, making RF suitable for noisy and high-dimensional data. It also provides
feature importance scores, but can be computationally intensive for large datasets
[36, 37].

• Multilayer Perceptron (MLP) is a type of neural network with multiple layers that
capture nonlinear relationships in data. Using activation functions such as ReLU or
sigmoid, MLPs can model complex patterns. However, they require careful regular-
ization to avoid overfitting and can be computationally demanding [38].



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 11 of 39

3.3.2. Homogeneous Ensemble Classifiers

Homogeneous ensembles combine multiple models of the same type to improve predictive
accuracy and reduce variance. These methods utilize the strengths of individual models to
enhance stability and reliability. The following are key homogeneous ensemble techniques
used in this study [28, 35].

• Bagging (Bootstrap Aggregating) creates diverse training sets by bootstrapping and
trains multiple instances of a base model. The final prediction is determined by
aggregating the output, typically using majority voting. Bagging is particularly
effective in reducing overfitting in models prone to high variance, such as decision
trees [27].

• AdaBoost (Adaptive Boosting) builds a sequence of models where each subsequent
model focuses on correcting errors made by its predecessors. The final predictions
are weighted on the basis of the models’ performance. AdaBoost improves weak
learners, but can be sensitive to noisy data and class imbalances [35].

• Gradient Boosting iteratively trains models to minimize residual errors from previous
models. The final output combines all the models’ predictions. This technique is
highly flexible, supports various data types, and provides insight into the importance
of features [27, 28].

• XGBoost (Extreme Gradient Boosting) improves traditional gradient boosting with
optimizations such as parallel processing and regularization, making it computation-
ally efficient and highly effective for complex datasets [39].

• Extra Trees generates an ensemble of extremely randomized decision trees. Intro-
duces randomness in feature selection and split point determination, improving gen-
eralization and efficiency [39, 40].

• CatBoost specializes in handling categorical data through efficient encoding tech-
niques. It includes mechanisms to prevent overfitting and supports GPU accelera-
tion, making it suitable for diverse datasets.

• Hist Gradient Boosting employs histogram-based decision trees to improve compu-
tational efficiency by approximating continuous values with histograms. It is partic-
ularly useful for large-scale problems [27, 35].

• Isolation Forests focus on identifying anomalies by isolating outliers using random
feature splits. Shorter isolation paths indicate potential anomalies, making this
method valuable for detecting rare patterns in high-dimensional data [41, 42].

3.3.3. Heterogeneous Ensemble Classifiers

Heterogeneous ensembles combine diverse classifiers to leverage their complementary strengths,
improving the accuracy and robustness of predictions. The key methods explored in this
study include stacking and voting, both of which are effective in detecting fake news.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 12 of 39

• Voting Classifiers aggregate predictions from different base models to create a unified
output. Hard voting relies on majority class votes, while soft voting uses weighted
probabilities based on model performance. Voting classifiers benefit from the diver-
sity of base models, which enhances overall predictive performance [34].

• Stacking Classifiers employ a meta-learner to combine predictions from multiple base
models. Unlike voting, stacking uses a data-driven approach to learn the optimal
combination of models output. Using the strengths of various models, stacking
often achieves superior performance compared to individual models or simple voting
schemes [34, 35].

3.4. Feature Selection Methods

Feature selection is a crucial step in machine learning that aims to identify the most
relevant features for model building. Effective feature selection not only improves model
performance but also reduces computational cost and helps prevent overfitting. In this
context, several techniques can be employed to select the most informative features. Three
prominent methods are Recursive Feature Elimination (RFE), SelectKBest, and Princi-
pal Component Analysis (PCA). Each of these methods utilizes distinct principles and
mathematical foundations to achieve feature selection.

3.4.1. Recursive Feature Elimination

RFE is an iterative feature selection technique that recursively removes the least significant
features based on the performance of a chosen model. The algorithm continues to remove
features until only the desired number of features remains. RFE works by fitting a model
to the data and ranking the features based on their importance. The features are then
removed on the basis of this ranking, and the model is re-trained on the reduced feature
set. This process is repeated until the number of features is equal to the target number,
k. RFE is typically used with models that provide feature importance scores, such as LR,
SVM, or DTs.

• Steps of RFE:

(i) Model Training: Train a model M on the initial set of features X to predict
the target y.

(ii) Feature Ranking: Evaluate the importance of each feature based on the
coefficients of the model or feature weights.

(iii) Feature Elimination: Remove the least important feature(s) from the feature
set.

(iv) Model Retraining: Re-train the model with the reduced set of features.

(v) Iteration: Repeat the process until the desired number of features k is reached.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 13 of 39

• Mathematical Formulation:

Let X be the matrix of features with samples n and features m, and y be the target
vector.

(i) Model Construction: Train a model M with features X:

f(X) = ŷ (2)

where ŷ is the predicted target vector.

(ii) Feature Importance: Compute the importance of each feature i:

Feature Importance = |wi| (3)

where wi is the coefficient for feature i.

(iii) Feature Elimination: Identify and remove the least important feature(s) based
on |wi|.

(iv) Update Feature Set: Update the feature set X and repeat the training and
elimination steps.

3.4.2. SelectKBest

SelectKBest is a univariate feature selection method that selects the top k features based
on univariate statistical tests. It evaluates each feature independently to determine its
relevance to the target variable. SelectKBest evaluates each feature using statistical tests
and selects the top k features based on test scores. Common statistical tests used include
the chi-square test, the ANOVA F-value, and mutual information.

• Steps of SelectKBest:

(i) Compute Test Statistics: Calculate the test statistic for each feature based
on a chosen statistical test.

(ii) Rank Features: Rank features according to their test statistics.

(iii) Select Top Features: Choose the best k features with the highest test statis-
tics.

• Mathematical Formulation:

Let X be the matrix of features with samples n and features m, and y be the target
vector.

(i) Compute Test Statistics: For each feature Xi, compute the test statistic Ti:

Ti = Test Statistic(Xi, y) (4)

(ii) Ranking Features: Rank features based on the computed test statistics:

Rank(Xi) = Sort(T1, T2, . . . , Tm) (5)

(iii) Selecting Top k: Select the k features with the highest Ti:

Xselected = Top k features based on Ti (6)



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 14 of 39

3.4.3. Principal Component Analysis

PCA is a dimensionality reduction technique that transforms the original feature space
into a new set of uncorrelated variables called principal components. PCA identifies the
directions (principal components) along which the variance of the data is maximized. PCA
transforms data into a new coordinate system where the first coordinate accounts for the
largest variance in the data, the second coordinate accounts for the next largest variance,
and so on. The goal is to reduce the number of dimensions while retaining as much
variance as possible.

• Steps of PCA:

(i) Compute Covariance Matrix: Calculate the covariance matrix of the fea-
ture set.

(ii) Eigen Decomposition: Perform eigenvalue decomposition to obtain eigen-
vectors and eigenvalues.

(iii) Sort Principal Components: Sort eigenvectors by the magnitude of their
corresponding eigenvalues.

(iv) Select Top Components: Choose the top k eigenvectors to form the principal
components.

(v) Project Data: Transform the original data into the new feature space spanned
by the selected principal components.

• Mathematical Formulation:

Let X be the feature matrix with n samples and m features.

(i) Compute Covariance Matrix:

Σ =
1

n− 1
XTX (7)

(ii) Perform Eigen Decomposition: Find eigenvectors W and eigenvalues Λ from
the covariance matrix:

ΣW = WΛ (8)

(iii) Sort Eigenvectors: Sort eigenvectors based on the eigenvalues in descending
order.

(iv) Select the top k components: Choose the top k eigenvectors to form the new
feature space.

(v) Project Data: Project the original data onto the new space:

Xnew = XW (9)



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 15 of 39

3.4.4. Genetic Algorithm

GAs simulate the process of natural evolution to solve optimization problems. They
operate on a population of potential solutions and use evolutionary processes such as
selection, crossover, and mutation to evolve better solutions over generations.

• Mathematical Formulation

The GA-based feature selection process involves the following components:

• Representation

Each individual in the GA population represents a subset of features using a binary
vector. If the i-th bit of the vector is 1, the corresponding feature is included in the
subset; otherwise, it is excluded. Let X be the feature matrix with n features and y
the target variable. Each individual x can be represented as:

x = [x1, x2, . . . , xn] (10)

where xi ∈ {0, 1}.

• Fitness Function

The fitness function evaluates the quality of the subset of features. In our case, the
fitness function is the accuracy of a classifier (Random Forest in this work) trained
on the selected features:

f(x) = Accuracy(Classifier(Xx, y)) (11)

where Xx denotes the feature matrix with features selected according to x.

• Genetic Algorithm Operations

The GA operations are as follows:

(i) Initialization: Generate an initial population of binary vectors.

(ii) Selection: Select individuals based on their fitness using tournament selection.

(iii) Crossover: Create new individuals by exchanging segments of binary vectors
between parents.

(iv) Mutation: Introduce variability by flipping bits in the binary vectors.

(v) Replacement: Form the next generation of the population.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 16 of 39

3.5. Performance Measures

In assessing the performance of the previously outlined classification ML methods for
the detection of fake news in tweets, it is crucial to evaluate their accuracy, recall, precision,
F1 score and AUC-ROC. These metrics provide valuable insights into the model’s ability
to correctly classify tweets as either real or fake, detect relevant instances, balance between
precision and recall, and evaluate the overall performance of the classification model. Given
the potential impact of false positives - where legitimate news can be incorrectly labeled
fake - it is particularly important to evaluate models not only based on overall accuracy,
but also in terms of their precision and recall. High precision helps minimize false positives,
which is crucial for maintaining trust and preventing the suppression of valid content. In
contrast, high recall ensures that actual instances of fake news are effectively identified.
The combination of these metrics supports a balanced and context-aware interpretation of
the behavior of the model, allowing stakeholders to prioritize detection objectives based
on specific deployment needs.

• Accuracy is a measure of how well a model is able to correctly classify tweets as real
or fake. It is calculated as the proportion of correctly classified instances (both real
and fake tweets) out of the total number of instances.

Accuracy =
TP + TN

TP + TN + FP + FN
(12)

where:

– TP (True Positives) refers to the number of fake tweets correctly classified as
fake.

– TN (True Negatives) refers to the number of real tweets correctly classified as
real.

– FP (False Positives) refers to the number of real tweets incorrectly classified
as fake.

– FN (False Negatives) refers to the number of fake tweets incorrectly classified
as real.

• Precision measures the model’s ability to avoid false positives, focusing on how many
of the tweets classified as fake are indeed fake. It is calculated as the proportion of
true positives out of the total number of instances classified as fake.

Precision =
TP

TP + FP
(13)

In the context of fake news detection in tweets, precision indicates the model’s ef-
fectiveness in identifying actual fake tweets among those it classifies as fake.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 17 of 39

• Recall measures the proportion of actual fake tweets that are correctly identified by
the model. It is calculated by dividing the number of true positives by the sum of
true positives and false negatives.

Recall =
TP

TP + FN
(14)

In the context of fake news detection in tweets, recall indicates the model’s ability
to detect all the fake tweets that are actually present.

• F1 Score is a harmonic mean of precision and recall, providing a balanced measure of
the model’s performance in detecting fake tweets. It accounts for both the precision
of the model’s predictions and its ability to capture all relevant instances.

F1 =
2 · Precision · Recall

Precision + Recall
(15)

The F1 score is particularly useful in the detection of fake news for tweets where
both false positives and false negatives have significant consequences, and the balance
between precision and recall is crucial.

• AUC-ROC (Area Under the Receiver Operating Characteristic Curve) is a perfor-
mance measurement for classification problems at various thresholds. It plots the
true positive rate (recall) against the false positive rate and measures the area under
this curve.

AUC-ROC =

∫ +∞

−∞
TPR(x) dFPR(x) (16)

where:

– TPR (True Positive Rate) is the recall of the model.

– FPR (False Positive Rate) is the proportion of actual negatives that are incor-
rectly classified as positives.

In the context of fake news detection in tweets, the AUC-ROC score provides an
aggregate measure of the model’s ability to discriminate between fake and real tweets
across all possible classification thresholds. A higher AUC-ROC score indicates a
better overall performance in distinguishing between fake and real tweets.

4. Results and Analysis

4.1. Individual Classifiers Results

Table 4 demonstrates the performance of various individual ML models in detecting
fake news using the TruthSeeker dataset. The results shown in this table reveal notable



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 18 of 39

insights into the effectiveness of these models in detecting fake news. Each model was
assessed based on five key metrics: accuracy, precision, recall, F1 score, and AUC-ROC,
which together provide a comprehensive view of their predictive capabilities. In this regard:

Table 4: Performance of Individual Classifiers on the TruthSeeker Dataset

Classifier Accuracy Precision Recall F1-score AUC-ROC

NB 92.969% 92.996% 92.969% 92.971% 93.414%

LR 95.801% 95.801% 95.801% 95.801% 95.889%
KNN 92.414% 92.414% 92.414% 92.414% 95.029%

MLP 95.387% 95.387% 95.387% 95.387% 95.847%

SVM 95.704% 95.704% 95.704% 95.704% 95.902%
DT 91.006% 91.007% 91.006% 91.006% 91.002%

RF 95.801% 95.801% 95.801% 95.801% 95.892%

• Naive Bayes: NB achieved an accuracy of 92.969%, with precision and recall close
to 93%, indicating its ability to consistently identify both true positives and true
negatives. Its AUC-ROC of 93.414% reflects a robust performance in distinguishing
between fake and real news.

• Logistic Regression: LR demonstrated superior performance with accuracy, pre-
cision, recall, and F1 score of 95.801%. The AUC-ROC score of 95.889% further
emphasizes its strong discriminatory power, making it one of the most reliable clas-
sifiers in this evaluation.

• K-Nearest Neighbors: KNN showed a slightly lower performance with an accu-
racy of 92.414%. Despite its consistent scores in precision, recall, and F1 score, its
AUC-ROC of 95.029% suggests that KNN can effectively handle the classification
task, although not as efficiently as some other methods.

• Multilayer Perceptron: MLP achieved an accuracy of 95.387%, with matching
precision, recall, and F1 score values. Its AUC-ROC of 95.847% indicates strong
predictive performance, making it a competitive choice for the detection of fake
news.

• Support Vector Machine: SVM stood out with an accuracy of 95.704% and
uniformly high scores for precision, recall, and F1 score. Its AUC-ROC of 95.902%
underscores its excellent ability to differentiate between classes, proving to be one
of the best performers.

• Decision Tree: DT showed a lower overall performance with an accuracy of 91.006%.
Its precision and recall values were slightly higher at 91.007%, but the F1 score and
AUC-ROC of 91.002% reflect its limitations in handling the data set as effectively
as other classifiers.

• Random Forest: RF matched LR in accuracy, precision, recall, and F1 score at 95.
801%. Its AUC-ROC of 95.892% confirms its reliability and robustness, showcasing
its potential as a powerful ensemble method for detecting fake news.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 19 of 39

In summary, LR, SVM and RF emerged as the leading classifiers, demonstrating high
precision and balanced performance across all metrics. These methods provide a solid
foundation for reliable fake news detection on the TruthSeeker dataset, highlighting their
practical applicability in real-world scenarios.

4.2. Ensemble Methods Results

Table 5 presents the performance metrics of various ensemble ML algorithms used for
the detection of fake news. These algorithms include both homogeneous and heterogeneous
ensemble methods, evaluated across five key metrics: Accuracy, precision, recall, F1 score,
and AUC-ROC. The results depicted in this table can be summarized as follows:

• CatBoost demonstrated the highest overall performance with an accuracy of 95.801%
and an equally high precision, recall, and F1 score of 95.801%. Its AUC-ROC score
of 95.859% highlights it as one of the main methods to distinguish between fake and
real news.

• HistGradient Boosting also achieved a high accuracy of 95.801% with consistent
metrics for precision, recall, and F1 score. Its AUC-ROC of 95.904% is the highest
among all methods tested, marking it as the most effective classifier for this dataset.

• LightGBM matched CatBoost and HistGradient Boosting with an accuracy of
95.801% and consistent Precision, Recall and F1 score. The AUC-ROC score of
95.854% further establishes it as a highly effective method for the detection of fake
news.

• XGBoost achieved an accuracy of 95.790% with an equivalent precision, recall, and
F1 score. Its AUC-ROC of 95.880% reflects strong performance and makes it a
competitive choice among the tested classifiers.

• Stacking Classifier produced an accuracy of 95.801% and matched the metrics for
Precision, Recall, and F1-score. With an AUC-ROC of 95.816%, it is also a strong
performer for the detection of fake news.

• Voting Classifier showed an accuracy of 95.786% with consistent Precision, Recall,
and F1-score values. The AUC-ROC of 95.892% indicates reliable performance,
although slightly less effective compared to the best methods.

• Bagging (Decision Tree) demonstrated high performance with an accuracy of
95.738% and uniform scores for precision, recall, and F1 score. Its AUC-ROC of
95.893% confirms its effectiveness in handling the classification task.

• Gradient Boosting provided strong results with an accuracy of 95.801% and con-
sistent metrics for precision, recall, and F1 score. Its AUC-ROC of 95.827% reflects
effective performance for fake news detection.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 20 of 39

• Extra Trees achieved an accuracy of 95.756% with very close values for precision,
recall, and F1 score. The AUC-ROC of 95.779% indicates strong performance, al-
though slightly lower than the top methods.

• AdaBoost (Decision Tree) showed a lower performance with an accuracy of
91.025% and consistent scores on all metrics. The AUC-ROC of 91.019% indicates
that it is less effective compared to other methods.

• Isolation Forest was the least effective with an accuracy of 51.140% and signifi-
cantly lower scores for precision, recall, and F1 score. The AUC-ROC of 47.959%
reflects its unsuitability for the detection of fake news, as it is designed for the
detection of anomalies rather than classification.

Table 5: Performance of Ensemble Classifiers on the TruthSeeker Dataset

Classifier Accuracy Precision Recall F1-score AUC-ROC

AdaBoost (Decision Tree) 91.025% 91.025% 91.025% 91.025% 91.019%
Bagging (Decision Tree) 95.738% 95.738% 95.738% 95.738% 95.893%

CatBoost 95.801% 95.801% 95.801% 95.801% 95.859%

Extra Trees 95.756% 95.757% 95.756% 95.756% 95.779%
Gradient Boosting 95.801% 95.801% 95.801% 95.801% 95.827%

HistGradient Boosting 95.801% 95.801% 95.801% 95.801% 95.904%

Isolation Forest 51.140% 26.376% 51.140% 34.802% 47.959%
LightGBM 95.801% 95.801% 95.801% 95.801% 95.854%

Stacking Classifier 95.801% 95.801% 95.801% 95.801% 95.816%

Voting Classifier 95.786% 95.786% 95.786% 95.786% 95.892%
XGBoost 95.790% 95.790% 95.790% 95.790% 95.880%

In summary, the best-performing methods for fake news detection are CatBoost, Hist-
Gradient Boosting, LightGBM, and XGBoost. Among these, HistGradient Boosting
stands out with the highest AUC-ROC score, while CatBoost and LightGBM also demon-
strate exceptional performance across all metrics. The Stacking Classifier and Voting
Classifier are also effective but slightly less so compared to the aforementioned top meth-
ods. Bagging (Decision Tree) and Gradient Boosting offer strong performance, though
they fall short of the best methods. Extra Trees and AdaBoost (Decision Tree) show solid
results but are not as competitive, while Isolation Forest proves to be inadequate for this
specific classification task.

Given that individual classifiers and ensemble methods demonstrated comparable lev-
els of performance and considering the reduced computational complexity associated with
individual models, the subsequent analysis focuses on evaluating the impact of feature
selection techniques on individual classifiers. In particular, feature selection methods such
as RFE and GA were applied exclusively to individual classifiers in this study, and not to
ensemble models. This methodological choice was made to isolate and assess the contri-
bution of feature selection strategies in improving model performance while maintaining
computational efficiency. The resulting analysis offers valuable information on the poten-
tial of optimized individual classifiers to deliver competitive performance with reduced



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 21 of 39

training overhead, thus supporting their applicability in resource-constrained or real-time
environments.

4.3. Impact of Feature Selection on Machine Learning Algorithms

Feature selection plays a crucial role in enhancing the performance and efficiency of ML
models. This section explores the impact of different feature selection techniques on the
performance of various classifiers in the context of fake news detection. Specifically, the
section investigates how the performance of ML algorithms is affected by selecting different
subsets of features and reducing dimensionality. The analysis begins by evaluating the
effectiveness of feature selection methods such as RFE and SelectKBest. These methods
are examined by varying the number of features from 5 to 30 and assessing the impact
on the performance metrics of classifiers including LR, NBs. Performance metrics such as
accuracy, precision, recall, F1 score, and AUC-ROC are used to gauge how the increasing
number of features influences the effectiveness of these models. Subsequently, PCA is
explored as a technique for dimensionality reduction. The performance of classifiers is
analyzed as the number of principal components is varied from 5 to 30, providing insights
into how dimensionality reduction affects model performance. Finally, the section covers
the use of GA for feature selection, focusing on how the optimal feature subset identified
by the GA affects the performance of ML algorithms. The GA approach is evaluated
on the basis of the optimal number of features selected and compared with the results
obtained from RFE, SelectKBest, and PCA. Through this examination, the purpose of
this section is to provide a comprehensive understanding of how different feature selection
strategies and feature dimensions impact the performance of ML algorithms.

4.3.1. Recursive Feature Elimination Results

Figure 1 shows the performance of various machine algorithms as the number of selected
features increases from 5 to 30 using the RFE feature selection method. For example,
Figure 1a illustrates the accuracy of the model as the number of features increases from 5
to 30. The accuracy results for different models using RFE exhibit various trends as the
number of features increases. LR maintains a high accuracy of 95.80% with 5 features,
which remains constant up to 10 features, but then decreases to 60.19% with 30 features.
NB shows a low and relatively constant accuracy in all feature sets, with 52.28% at 5
features and 49.88% at 30 features. KNN starts with a higher accuracy of 72.84% at five
features and then stabilizes at 58.73% from 10 features onward. SVM shows a consistently
low accuracy of around 53.49% in all feature sets, indicating a minimal sensitivity to the
number of features. MLP starts with a high accuracy of 95.73% at 5 features, decreases
to 79.09% at 10 features, and fluctuates before increasing to 86.53% at 30 features. DT
exhibits high and stable accuracy, ranging from 91.36% at 5 features to 90.80% at 30
features. Finally, RF maintains a high and consistent accuracy from 95.70% at 5 features
to 95.80% at 30 features, demonstrating the most stable performance. On the other hand,
figure 1b shows precision values as the number of features increases.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 22 of 39

5 10 15 20 25 30
Number of Features

45

50

55

60

65

70

75

80

85

90

95

100

A
cc

ur
ac

y 
(%

) LR
NB
KNN
SVM
MLP
DT
RF

a Accuracy

5 10 15 20 25 30
Number of Features

45

50

55

60

65

70

75

80

85

90

95

100

P
re

ci
si

on
 (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

b Precision

5 10 15 20 25 30
Number of Features

45

50

55

60

65

70

75

80

85

90

95

100

R
ec

al
l (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

c Recall

5 10 15 20 25 30
Number of Features

30

40

50

60

70

80

90

100

F
1 

S
co

re
 (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

d F1 Score

5 10 15 20 25 30
Number of Features

50

55

60

65

70

75

80

85

90

95

100

A
U

C
-R

O
C

 (
%

) LR
NB
KNN
SVM
MLP
DT
RF

e AUC-ROC

Figure 1: Performance Metrics of Various Models as Feature Number Increases under
Recursive Feature Elimination.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 23 of 39

The precision results for different models using RFE show a varied performance in
terms of number of features. For LR, the precision starts at 95.80% with 5 features and
decreases to 60.34% as the number of features increases to 30. NB shows a relatively
stable precision, with 58.89% at 5 features and 57.08% at 30 features. KNN maintains a
precision level from 72.83% at 5 features to 58.72% at 30 features, indicating a gradual
decrease. The SVM has stable but low precision, starting at 53.93% and slightly increasing
to 54.78% as more features are used.

MLP begins with a high precision of 95.73% at 5 features, which drops to 81.86% at 10
features, and then fluctuates before increasing to 88.14% at 30 features. DT shows high
and consistent precision throughout, with values from 91.37% to 90.80% throughout the
range of features. RF exhibits high and steady precision across all set of features, starting
at 95.70% and remaining at 95.80%.

The recall results for different models using RFE display diverse trends as the number
of features increases as shown in figure 1c. LR achieves a high recall of 95.80% in both 5
and 10 features, but experiences a significant decline to 60. 19% in 30 features. NB shows
a consistently low recall, with values of 52.28% at five features and remaining around
49.88% at 30 characteristics. KNN begins with a high recall of 72.84% for five features
and stabilizes at 58.73% for feature counts of 10 and above. SVM maintains a constant
recall of 53.49% across all feature sets, indicating that it is relatively unaffected by the
number of features. MLP starts with a high recall of 95.73% at 5 features, decreases to
79.09% at 10 features, and then fluctuates before rising to 86.53% at 30 features, showing
an overall trend of improvement with more features. DT shows a high and stable recall
from 91.36% at 5 features to 90.80% at 30 features, demonstrating robust performance
regardless of the number of features. Lastly, RF also shows high and consistent recall
from 95.70% with 5 features to 95.80% with 30 features, representing the most reliable
and effective model in terms of recall. The results of the F1 score for various models
using RFE reveal different trends as the number of features increases as depicted in figure
1d. LR maintains a high F1 score of 95.80% for 5 and 10 features, but shows a notable
decrease to 59.75% for 30 features, indicating a decrease in performance as more features
are added. NB exhibits low F1 scores throughout, starting at 43.73% with 5 features
and stabilizing around 36.80% from 10 to 30 features, reflecting poor performance with
increasing features. KNN achieves a high F1 score of 72.82% for five features, but then
shows a decrease to about 58.73% for feature counts of 10 and above. SVM demonstrates a
relatively consistent F1 score of 51.72% in most feature sets, with a slight peak at 53.94%
with 5 features. MLP starts with a high F1 score of 95.73% at 5 features, drops to 78.74%
at 10 features, and fluctuates before reaching 86.43% at 30 features, suggesting improved
performance with more features. DT shows high and stable F1 scores from 91.36% at 5
features to 90.80% at 30 features. Lastly, RF maintains high and consistent F1 scores from
95.70% at 5 features to 95.80% at 30 features, reflecting the most effective and reliable
performance in terms of the F1 score.

The AUC-ROC results for various models using RFE illustrate different performance
trends as the number of features increases, as illustrated in Figure 1e. LR and MLP
both achieve high AUC-ROC scores of 95.75% at 5 features. However, while LR exhibits



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 24 of 39

a decrease in AUC-ROC to 64.44% by 30 features, MLP maintains a high AUC-ROC
with scores of 89.78% at 10 features and 89.78% at 30 features. NB shows a significant
drop in AUC-ROC from 87.87% at 5 features to 58.24% at 30 features. KNN achieves
a maximum AUC-ROC of 79.51% at 5 features but stabilizes around 62.61% from 10
features onward. SVM consistently shows a low AUC-ROC score of 55.73% in all feature
sets. DT demonstrates stable AUC-ROC scores of approximately 90.80% on all feature
counts. Finally, RF maintains the highest and most consistent AUC-ROC scores, ranging
from 95.81% at 5 features to 95.85% at 30 features.

4.3.2. SelectKBest Restlts

Figure 2 illustrates the impact of feature selection on the performance of various ML
algorithms as the number of selected features increases from 5 to 30. As shown in the
figure 2a, the accuracy of various ML models is affected by increasing the number of
features selected using the SelectKBest method from 5 to 30. For LR, DT, and RF, the
accuracy remains relatively stable across the range of features. Specifically, LR maintains
a high accuracy of approximately 95.80% throughout the feature selection process. SVM
exhibits a significant decrease in accuracy, dropping from 95.80% at 5 features to 53.99%
at 30 features, indicating a considerable decline in performance as the number of features
increases. In contrast, NB and KNN also show a decrease in accuracy, but not as drastic
as SVM, with NB falling from 95.80% at 5 features to around 61.20% at 30 features and
KNN decreasing from 95.69% at 5 features to 57.34% at 30 features. MLP demonstrates
variable performance, with a notable drop in 15 features but improved to around 90.64% at
30 features. Overall, most models achieve optimal or near-optimal accuracy with a higher
number of features, except for SVM, which experiences a significant drop in performance.
Figure 2b illustrates how the precision of various ML models changes as the number of
features selected using the SelectKBest method increases from 5 to 30. For LR, DT,
and RF, the precision remains relatively high and stable throughout the range of feature
selection. LR maintains a consistent precision of approximately 95.80%, and RF also
shows high precision around 95.80% throughout the feature selection process. In contrast,
NB exhibits a significant decline in precision from 95.80% at 5 features to approximately
61.22% at 30 features, with only minor fluctuations at intermediate feature counts. KNN
also experiences a decrease in precision from 95.69% with 5 features to 57.32% with 30
features. The MLP shows a fluctuating precision, dropping to 78.46% at 15 features
but recovering to around 90.84% at 30 features. The DT maintains relatively constant
precision around 90.63% to 90.69% throughout the feature range. Overall, most models see
stable or improved precision as the number of features increases, with notable exceptions
for NB and KNN, which exhibit considerable variability.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 25 of 39

5 10 15 20 25 30
Number of Features

50

60

70

80

90

100

A
cc

ur
ac

y 
(%

)

LR
NB
KNN
SVM
MLP
DT
RF

a Accuracy

5 10 15 20 25 30
Number of Features

50

60

70

80

90

100

P
re

ci
si

on
 (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

b Precision

5 10 15 20 25 30
Number of Features

50

60

70

80

90

100

R
ec

al
l (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

c Recall

5 10 15 20 25 30
Number of Features

50

60

70

80

90

100

F
-1

 S
co

re
 (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

d F1 Score

5 10 15 20 25 30
Number of Features

50

60

70

80

90

100

A
U

C
-R

O
C

 (
%

)

LR
NB
KNN
SVM
MLP
DT
RF

e AUC-ROC

Figure 2: Performance Metrics of Various Models as Feature Number Increases under
SelectKBest.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 26 of 39

Figure 2c illustrates how the recall of various ML models is affected by increasing the
number of features selected using the SelectKBest method from 5 to 30. For LR, MLP,
DT, and RF, recall remains relatively high and stable throughout the range of features.
LR and RF maintain high recall values, remaining above 95.66% for all feature counts. DT
also shows stable high recall around 90.63% to 90.69%. In contrast, NB and KNN exhibit
significant variability in recall. NB starts with a high recall of 95.80% at 5 features but
drops to 61.20% at 30 features, with a slight increase at 30 features compared to 15 and
25 features. KNN also shows a decrease in recall from 95.69% at 5 features to 57.34% at
30 features, reflecting a steady decline in performance. SVM shows a generally low recall
across all feature counts, starting at 95.80% at 5 features and decreasing to 53.99% at 30
features. MLP exhibits a notable increase in recall from 68.71% at 15 features to 90.64%
at 30 features. In general, most models maintain or improve recall with more features,
except for NB, KNN, and SVM, which show significant declines in recall performance.

Figure 2d shows the F-1 scores of various ML models as the number of features selected
using the SelectKBest method increases from 5 to 30. LR maintains a consistently high
F-1 score close to 95.80% under all feature counts. RF also shows a stable high F-1
score around 95.80%, demonstrating consistent performance regardless of the number of
features. NB and KNN exhibit notable decreases in F-1 score as the number of features
increases. NB starts with a high F-1 score of 95.80% at five features but decreases to
around 61.02% at 30 features. Similarly, KNN shows a drop in the F-1 score from 95.69%
at 5 features to 57.32% at 30 features, reflecting a steady decline in performance. SVM
maintains a low and relatively stable F-1 score throughout the feature range, starting at
95.80% at 5 features and dropping to 54.00% at 30 features, showing a significant decline
in performance as the number of features increases. MLP exhibits fluctuations in the F-1
score, with a significant drop to 65.38% at 15 features but recovered to approximately
90.62% at 30 features. DT maintains a relatively stable F-1 score of around 90.63% to
90.69% throughout the feature selection process. In general, most models, except NB,
KNN and SVM, maintain or improve their F-1 scores as the number of features increases.
NB and KNN show significant decreases, while SVM’s performance declines substantially.

Figure 2e presents the AUC-ROC scores for different ML models as the number of
features selected using the SelectKBest method increases from 5 to 30. LR maintains a
high and relatively stable AUC-ROC score, starting at 95.83% with 5 features and slightly
decreasing to 95.65% with 30 features. Similarly, RF exhibits a high and stable AUC-
ROC score, remaining between 95.58% and 95.74% in the range of features. NB and KNN
demonstrate a consistent decline in AUC-ROC as more features are added. NB has an
initial AUC-ROC of 95.82% at 5 features, but this score drops to 65.82% at 30 features.
KNN also shows a decrease from 95.77% at 5 features to 60.99% at 30 features, reflecting a
clear downward trend in performance. SVM maintains a lower and more stable AUC-ROC
score across all feature counts, starting at 95.78% at 5 features and decreasing marginally
to 55.56% at 30 features. This indicates a steady decline in SVM’s performance as more
features are considered. MLP shows significant variability in AUC-ROC, with a marked
drop to 77.41% at 15 features but recovered to 92.15% at 30 features, showing that the
performance of this model is sensitive to the number of features selected. DT maintains



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 27 of 39

a relatively stable AUC-ROC score between 90.62% and 90.93% throughout the feature
selection process, indicating consistent performance with the increasing number of features.
Most models other than NB and KNN exhibit stable or improved AUC-ROC scores as the
number of features increases. The performance of NB and KNN deteriorates significantly,
while SVM shows a consistent decline, and MLP exhibits fluctuating performance.

Evaluation of ML models using four different performance metrics: accuracy, precision,
recall, and AUC-ROC reveals distinct patterns in how these models perform as the number
of features increases from 5 to 30 using the SelectKBest method. LR and RF exhibit robust
and consistent performance across all metrics. For accuracy and AUC-ROC, LR remains
stable at approximately 95.80% and 95.83% respectively, while RF also maintains high
performance with accuracy around 95.80% and AUC-ROC between 95.58% and 95.74%.
MLP shows variability in performance, with fluctuating precision, recall, and AUC-ROC,
reflecting the sensitivity to the number of features selected.

In contrast, NB and KNN show significant declines in performance metrics as the
number of features increases. NB experiences substantial drops in precision, recall, and
AUC-ROC from high initial values to around 61.02%, 65.82%, and 65.82% respectively, at
30 features. Similarly, KNN demonstrates a decrease in accuracy from 95.69% to 57.32%,
precision from 95.69% to 57.32%, and AUC-ROC from 95.77% to 60.99% over the same
range. SVM consistently shows lower performance across all metrics, with a substantial
decrease in recall and AUC-ROC as the number of features increases, and a lower AUC-
ROC value from 95.78% to 55.56%. Finally, DT exhibits stable performance in accuracy,
precision and recall, maintaining a consistent AUC-ROC around 90.62% to 90.93%.

4.3.3. Principal Component Analysis Results

This section shows the results of applying PCA to reduce dataset dimensionality and its
effects on the performance of various ML algorithms. Examines how increasing the num-
ber of principal components from 5 to 30 influences the accuracy, precision, recall, and
AUC-ROC of the models. Figure 3a illustrates the accuracy of various ML models as the
number of principal components increases from 5 to 30. The accuracy results obtained
through PCA are generally lower compared to those of other feature selection methods. RF
achieves the highest accuracy, with an increase from 61.34% at 5 components to 64.59% at
30 components, demonstrating the best performance among the models evaluated. In con-
trast, MLP exhibits the lowest accuracy, starting at 49.23% at 5 components and slightly
improving to 52.40% at 30 components. KNN and SVM maintain relatively stable accu-
racies around 58.67% and 53.50%, respectively, while LR shows a gradual increase from
54.17% to 58.30%. DT demonstrates relatively lower performance, with accuracy values
ranging from 58.58% to 56.07%. Overall, the use of PCA for dimensionality reduction
results in lower accuracy compared to other feature selection methods, indicating that
while PCA may reduce dimensionality, it does not necessarily enhance model performance
as effectively as methods like RFE or SelectKBest.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 28 of 39

5 10 15 20 25 30
Number of Features

50

55

60

65

A
cc

ur
ac

y 
(%

)

LR
NB
KNN
SVM
MLP
DT
RF

a Accuracy

5 10 15 20 25 30
Number of Features

45

50

55

60

65

P
re

ci
si

on
 (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

b Precision

5 10 15 20 25 30
Number of Features

50

55

60

65

R
ec

al
l (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

c Recall

5 10 15 20 25 30
Number of Features

35

40

45

50

55

60

65

F
-1

 S
co

re
 (

%
)

LR
NB
KNN
SVM
MLP
DT
RF

d F1 Score

5 10 15 20 25 30
Number of Features

45

50

55

60

65

70

75

A
U

C
-R

O
C

 (
%

)

LR
NB
KNN
SVM
MLP
DT
RF

e AUC-ROC

Figure 3: Performance Metrics of Various Models under PCA.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 29 of 39

Figure 3b illustrates the precision of various ML models as the number of principal
components increases from 5 to 30. The precision results for PCA are generally low com-
pared to other feature selection methods, highlighting a notable deficiency in performance
for most models. RF achieves the highest precision, increasing from 61.34% at 5 compo-
nents to 64.60% at 30 components, indicating that it is the most effective model in terms
of precision under PCA. In contrast, MLP exhibits the lowest precision, with a substantial
drop from 47.77% at 5 components to 52.93% at 30 components, which shows that this
model performs poorly in terms of precision. KNN and SVM maintain relatively stable
precision values around 54.78% and 58.73%, respectively. LR and NB show a general in-
crease in precision as the number of components grows, though their performance remains
lower than that of RF. Specifically, LR improves from 54.09% to 58.30%, and NB shows
a slight increase from 55.24% to 56.26%. Overall, the precision results underscore that
PCA’s dimensionality reduction approach leads to relatively poor performance across the
models, with RF being the exception and MLP showing significant performance issues.
Figure 3c shows the recall values for different MLP models as the number of principal
components increases from 5 to 30. The recall values, which represent the proportion of
true positive instances identified by each model, reveal a generally low performance across
the models compared to other feature selection methods. RF exhibits the highest recall,
increasing from 61.34% at 5 components to 64.59% at 30 components. This indicates
that RF performs the best in identifying true positive instances as the number of prin-
cipal components grows. KNN shows stable recall values, remaining around 58.67% to
58.73% throughout the range of principal components, reflecting consistent performance.
In contrast, MLP has relatively poor recall performance, with values starting at 49.23%
at 5 components and increasing to 52.40% at 30 components. LR and NB demonstrate
moderate recall improvements, with LR increasing from 54.17% to 58.30% and NB in-
creasing from 49.76% to 52.01%. SVM maintains low and stable recall values, ranging
from 53.49% at 5 components to 53.49% at 30 components, showing minimal improve-
ment as the number of components increases. DT shows a gradual increase in recall from
58.58% to 56.07%, reflecting its generally strong performance compared to most models.
In general, the recall values are relatively low in all models compared to other feature
selection methods, indicating that PCA does not significantly improve the ability of these
models to identify true positives. Random Forest is the most effective in terms of recall,
whereas other models such as SVM and MLP show relatively poor performance.

Figure 3d illustrates the F-1 scores for various ML models as the number of principal
components increases from 5 to 30. The F-1 score, which balances precision and recall,
highlights the trade-offs between these two metrics. RF consistently achieves the highest
F-1 scores in all principal components, starting at 61.34% with 5 components and increas-
ing to 64.59% with 30 components. This model demonstrates high precision and recall,
resulting in the most balanced performance among the models tested. In contrast, KNN
shows a stable F-1 score, remaining between 58.67% and 58.74%, indicating consistent
performance but at a lower level compared to Random Forest. DT exhibits moderate per-
formance in terms of the F-1 score, with values starting at 58.57% at 5 components and
reaching 56.07% at 30 components, showing a slight decrease as the number of principal



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 30 of 39

components increases. LR shows a steady improvement from 54.03% to 58.30%, reflecting
a positive trend in its ability to balance precision and recall. MLP shows relatively poor
F-1 scores, with values starting at 44.65% and increasing to 45.33%, indicating that this
model struggles to effectively balance precision and recall. NB also shows low F-1 scores,
starting at 37.21% and increasing to 45.04%, with a notable but still limited improvement
in performance. Finally, SVM maintains low and relatively stable F-1 scores throughout
the range of principal components, starting at 51.74% at 5 components and remaining
nearly unchanged up to 30 components. This reflects a consistent, yet underwhelming,
performance in achieving a balance between precision and recall. In summary, the F-1
score analysis reveals that while RF leads in balanced performance, other models such as
KNN, LR, and DT show varying levels of effectiveness, and MLP and NB demonstrate rel-
atively poor performance. The use of PCA for feature selection does not lead to significant
improvements in F-1 scores in most models.

Figure 3e displays the AUC-ROC scores for various ML models as the number of prin-
cipal components increases from 5 to 30. The AUC-ROC score measures the performance
of a model in distinguishing between classes, with higher values indicating better per-
formance. RF demonstrates the highest AUC-ROC scores among all models, starting at
68.10% with 5 principal components and increasing to 70.18% with 30 components. This
indicates that RF consistently performs well in terms of class separation as more princi-
pal components are considered. KNN also performs well in AUC-ROC, showing steady
and high performance from 62.60% at 5 components to 62.60% at 30 components. This
indicates that KNN maintains strong performance throughout the feature selection pro-
cess. LR and NB show moderate AUC-ROC scores. LR starts at 55.95% and increases to
61.55%, reflecting a trend of improved performance with more principal components. NB
starts at 55.04% and increases to 58.38%, showing a more modest improvement in perfor-
mance. MLP shows relatively high AUC-ROC values, starting at 48.11% and increasing to
58.47% as the number of principal components increases. This shows a gradual improve-
ment in performance, though it remains lower compared to the top-performing models. DT
exhibits the lowest AUC-ROC scores among the models, with values ranging from 54.45%
to 56.04%. This indicates that DT struggles more than other models in distinguishing
between classes effectively. SVM maintains relatively stable but low AUC-ROC perfor-
mance, starting at 55.73% and showing minimal variation up to 30 principal components.
This indicates that SVM’s ability to distinguish between classes is limited compared to
other models. In summary, RF leads in AUC-ROC performance, showing the best results
among all models as the number of principal components increases. KNN also performs
well, while other models like DT and SVM show relatively poor AUC-ROC scores. Most
models exhibit some improvement with more principal components, but overall perfor-
mance is modest compared to other feature selection methods.

Principal Component Analysis (PCA) consistently yielded lower performance com-
pared to other feature selection methods in all classifiers. This underperformance can
theoretically be explained by the nature of PCA itself. As an unsupervised dimensionality
reduction technique, PCA transforms the original feature space into orthogonal compo-
nents based on variance, without considering class labels. Although this helps to capture



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 31 of 39

Table 6: Selected Features by the Genetic Algorithm (GA)

No. Selected Feature Category

1 BinaryNumTarget Meta-data

2 statuses count Meta-data

3 following Meta-data
4 BotScoreBinary Meta-data

5 normalize influence Meta-data

6 mentions Meta-data
7 replies Meta-data

8 favourites Meta-data

9 ORG percentage Text
10 DATE percentage Text

11 FAC percentage Text
12 LAW percentage Text

13 PRODUCT percentage Text

14 LOC percentage Text
15 LANGUAGE percentage Text

16 Word count Text

17 Min word length Text
18 present verbs Lexical

19 past verbs Lexical

20 adverbs Lexical
21 questions Lexical

22 ampersand Lexical

23 long word freq Lexical

the overall structure of the data, it does not necessarily preserve the most relevant features
for distinguishing between real and fake content. In contrast, methods such as RFE and
SelectKBest directly evaluate the relevance of each feature with respect to the classifica-
tion target, making them more effective in contexts where feature-label relationships are
critical.

4.3.4. Genetic Algorithm Results

This section presents the results obtained by various ML models, following a feature
selection step performed using the GA. The GA algorithm optimally selected 23 features, as
shown in Table 6. This set of features was used to train and evaluate multiple ML models,
and their performance was assessed based on metrics including accuracy, precision, recall,
F1 score, and AUC-ROC, as shown in figure 4. To minimize the risk of overfitting during
the selection of features based on genetic algorithms, a 5-fold cross-validation strategy
was integrated into the GA fitness evaluation. Following the feature selection process, all
classifiers were re-trained and evaluated on a separate unseen test set to ensure robust and
unbiased performance assessment.

Figure 4a illustrates the accuracy scores of various ML models following the feature
selection step performed by the GA algorithm. The DT and RF models exhibit the highest
accuracy among all models, achieving nearly 91% and 96%, respectively. LR also shows
a high accuracy of approximately 96%, indicating strong performance on the test set. In
contrast, the KNN and SVM models demonstrate significantly lower accuracy, with values
of around 56% and 54%, respectively.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 32 of 39

a Accuracy b Precision

c Recall d F1 Score

e AUC-ROC

Figure 4: Performance Metrics of Various Models under GA Feature Selection.

MLP performs moderately with an accuracy of approximately 70%, while NB shows
the lowest accuracy at approximately 52%. Overall, the accuracy results highlight that
ensemble methods such as DT and RF perform significantly better compared to other
models following the feature selection step. Figure 4b presents the precision scores for



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 33 of 39

different models after applying the GA feature selection algorithm. LR and RF achieve
the highest precision scores of approximately 96% and 96%, respectively, indicating that
these models are effective at correctly identifying positive instances. The DT model also
shows high precision, around 91%, demonstrating its effectiveness in distinguishing positive
cases. MLP shows a good precision of about 80%. In contrast, the KNN and SVM models
exhibit lower precision scores, around 56% and 54%, respectively. The NB model achieves
a moderate precision score of approximately 58%. This figure underscores that LR, RF,
and DT are highly precise, while other models like KNN and SVM show relatively poor
performance in terms of precision. Figure 4c shows the recall scores for various models
after feature selection using the GA algorithm. LR and RF achieve the highest recall
scores, reaching around 96% and 96%, respectively. DT follows closely with a recall score
of approximately 91%, indicating an effective identification of positive instances. MLP
shows a moderate recall score of about 70%, while KNN and SVM exhibit significantly
lower recall scores, approximately 56% and 54%, respectively. NB has the lowest recall
score of around 52%. Recall analysis reveals that LR and RF are the most effective in
detecting positive instances, while other models such as KNN and SVM perform poorly
in this regard.

Figure 4d displays the F1 scores for different ML models after the GA feature selection
process. The LR and RF models lead with the highest F1 scores, approximately 96% and
96%, respectively. DT also shows a high F1 score of about 91%, demonstrating balanced
performance between precision and recall. MLP exhibits a moderate F1 score of around
68%. In contrast, the KNN and SVM models show relatively low F1 scores, about 56%
and 54%, respectively. NB achieves the lowest F1 score of approximately 36%. This figure
highlights that the ensemble methods, specifically DT and RF, along with LR, demonstrate
superior overall performance compared to other models. Figure 4e presents the AUC-ROC
scores for different models following the GA feature selection process. RF achieves the
highest AUC-ROC score of approximately 95.76%, demonstrating the best performance
in distinguishing between positive and negative instances. LR follows closely with a score
of around 95.85%, which also indicates strong performance. On the other hand, KNN and
SVM exhibit lower AUC-ROC scores of about 58.99% and 55.53%, respectively, reflecting
their relatively weaker performance in class discrimination. NB achieves a moderate AUC-
ROC score of approximately 69.07%, while DT shows an AUC-ROC score of approximately
90.72%, which is lower compared to LR and RF but still reflects reasonable performance.
In general, the results highlight that LR and RF demonstrate the highest effectiveness in
classifying instances, whereas KNN and SVM show relatively poor performance.

4.4. Summary of key insights

The evaluation of feature selection methods—RFE, SelectKBest, PCA, and GA—reveals
several important trends in how these techniques influence model performance. This sec-
tion discusses the general insights gained from these methods and highlights the stability
of different ML models across various feature selection techniques.

In general, the findings of this study can be summarized as the following key insights.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 34 of 39

• Performance Stability Across Feature Selection Methods: RF and LR are
notable for their stable performance across different feature selection methods. RF,
in particular, shows high and consistent performance across RFE, SelectKBest, and
GA methods. For example, RF achieves the highest AUC-ROC score of 95.85%
under the GA method, demonstrating its robustness in handling different subsets
of features. However, while LR maintains high performance under SelectKBest,
it is less consistent under RFE and PCA, indicating that its effectiveness is more
dependent on the feature selection process.

• Effectiveness of Dimensionality Reduction Techniques: The PCA technique
generally proves less effective compared to other feature selection methods. PCA’s
dimensionality reduction often leads to decreased model performance due to the loss
of important feature information and interpretability. This is in contrast to RFE
and SelectKBest, which focus on selecting the most relevant features and thereby
improve model performance. In this study, PCA fails to effectively enhance model
performance, reinforcing the idea that PCA is less suitable for tasks that require
feature relevance and interpretability.

While Principal Component Analysis is well regarded for its dimensionality reduc-
tion capabilities, it inherently compromises feature interpretability, which can hinder
explainability in applications such as fake news detection. In this study, PCA was
intentionally included as a benchmarking tool to examine the trade-offs between
dimensionality reduction and interpretability. Although PCA contributed to perfor-
mance improvement in limited cases, it was consistently outperformed by RFE and
GA, which not only yielded better classification results, but also preserved trans-
parency at the feature level. These findings reinforce the practical advantages of
using interpretable feature selection techniques for fake news detection.

• Success of Genetic Algorithms : GA emerges as a highly effective feature selec-
tion technique. The ability of GA to explore a wide range of feature combinations
allows it to identify optimal subsets that significantly enhance the performance of
the model. This effectiveness is demonstrated by RF achieving the highest AUC-
ROC score of 95.85% with the GA method. GA’s success highlights its potential as
a powerful tool for feature selection, providing a robust approach to optimize feature
subsets and improve classification results.

In summary, this study reveals that different feature selection methods have differ-
ent impacts on model performance. RFE and SelectKBest demonstrate effectiveness in
improving model outcomes, whereas PCA tends to degrade performance due to reduced
feature relevance and interpretability. Among the models evaluated, Random Forest (RF)
consistently shows high reliability and strong performance across all feature selection tech-
niques, highlighting its robustness. Logistic regression (LR) also performs competitively,
but exhibits greater sensitivity to the selected feature selection method. Genetic Algo-
rithms (GAs) further prove effective in optimizing feature subsets, as evidenced by the



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 35 of 39

superior results obtained when combined with RF. These findings emphasize the impor-
tance of pairing strong classifiers with interpretable and computationally efficient feature
selection strategies for fake news detection tasks. Furthermore, the evolving nature of fake
news poses a significant challenge to static models, particularly in the face of adversarial
content designed to bypass traditional detection mechanisms. Although the current ap-
proach relies on offline learning, these observations point to the need to explore incremental
and adaptive machine learning methods that can dynamically adjust to new patterns and
threats. Such approaches have the potential to improve the long-term effectiveness and
resilience of fake news detection systems.

In addition to technical considerations, ethical implications also play a crucial role in
the development and deployment of fake news detection systems. A major concern is the
possibility of false positives, cases in which true information is incorrectly flagged as fake,
which can suppress valid discourse, damage reputations, and undermine public trust. To
address this, the study incorporated a rigorous evaluation framework using a comprehen-
sive set of performance metrics, including precision, recall, F1 score, and AUC-ROC, to
ensure a balanced understanding of model behavior. However, technical accuracy alone
is not sufficient. Responsible deployment should consider transparency, accountability,
and, where appropriate, human oversight. Future work could benefit from incorporating
fairness-aware techniques and ethical design principles to ensure that detection systems
are not only accurate but also aligned with societal values.

5. Conclusions and Future Work

This study provided a comprehensive analysis of various machine learning classifiers
and feature selection techniques to enhance the detection of fake news. Multiple models,
including RF, Gradient Boosting, and SVM, were evaluated to identify the strengths and
limitations of each approach in the context of detecting fake news. The findings indicated
that ensemble learning methods, particularly RF and Gradient Boosting, achieved supe-
rior performance in terms of accuracy and robustness. This improvement was especially
pronounced when these models were paired with effective feature selection techniques, such
as RFE and SelectKBest, which reduced dimensionality and enhanced model performance
by focusing on the most relevant features.

The impact of different feature selection methods on the effectiveness of various classi-
fiers was thoroughly examined, demonstrating that techniques like RFE and SelectKBest
significantly improved performance. In contrast, PCA was found to be less effective, un-
derscoring the importance of carefully selecting feature extraction methods tailored to the
specific application.

In conclusion, the integration of robust ML models with optimized feature selection
strategies was shown to be critical to achieving high accuracy and reliability in fake news
detection. This study provided valuable information on the development of more effective
fake news detection systems and contributed to the advancement of methodologies in this
field.

While this study demonstrated the effectiveness of various traditional and ensemble



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 36 of 39

classifiers combined with feature selection techniques, future research could explore ad-
ditional avenues to further enhance the detection of fake news. One promising direction
involves the investigation of incremental and online learning approaches that allow models
to continuously adapt to new data. Given the dynamic and evolving nature of misinfor-
mation, such adaptive techniques can improve response to emerging trends without the
need for complete retraining. Additionally, evaluating model performance across multiple
datasets from different domains and platforms could help assess generalizability. Incorpo-
rating explainable AI (XAI) techniques can also strengthen the transparency and trustwor-
thiness of detection systems, particularly in high-stakes information environments. These
directions aim to build on the current work and extend its applicability to real-world,
evolving contexts.

Another potential extension is the inclusion of a comparative analysis of the com-
putational cost associated with different feature selection techniques in various types of
classifiers. Although the present study focused primarily on classification performance
and given that the work does not target deployment in resource-constrained environments
such as IoT or edge devices, computational efficiency was not emphasized. However, such
analysis would be highly valuable in scenarios that require trade-offs between accuracy
and efficiency, and could inform deployment decisions in constrained environments.

Acknowledgements

Machine learning training and evaluation have been performed using the Phoenix High
Performance Computing facility at the American University of the Middle East, Kuwait.

References

[1] Femi Olan, Uchitha Jayawickrama, Emmanuel Ogiemwonyi Arakpogun, Jana Suklan,
and Shaofeng Liu. Fake news on social media: the impact on society. Information
Systems Frontiers, 26(2):443–458, Apr 2024.

[2] H. Allcott and M. Gentzkow. Social media and fake news in the 2016 election. Journal
of Economic Perspectives, 31(2):211–236, 2017.

[3] A. Gupta and A. Roy. Manipulation of social media during the 2019 indian general
elections. Asian Journal of Communication, 29(5):537–556, 2019.

[4] M. Cinelli and A. Galeazzi. The covid-19 social media infodemic. Scientific Reports,
10:1–10, 2020.

[5] Omar Abdulwahabe Mohamad Rasha Talal Hameed. Federated learning in iot: A
survey on distributed decision making. Babylonian Journal of Internet of Things,
2023:1–7, Jan. 2023.

[6] Dipti Theng and Kishor K. Bhoyar. Feature selection techniques for machine learning:
a survey of more than two decades of research. Knowledge and Information Systems,
66(3):1575–1637, Mar 2024.

[7] I. Guyon and A. Elisseeff. An introduction to variable and feature selection. Journal
of Machine Learning Research, 3:1157–1182, 2003.



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 37 of 39

[8] Girish Chandrashekar and Ferat Sahin. A survey on feature selection methods. Com-
puters & Electrical Engineering, 40(1):16–28, 2014.

[9] Michael Greenacre, Patrick J. F. Groenen, Trevor Hastie, Alfonso Iodice D’Enza,
Angelos Markos, and Elena Tuzhilina. Principal component analysis. Nature Reviews
Methods Primers, 2(1):100, Dec 2022.

[10] Rania Saidi, Waad Bouaguel Ncir, and Nadia Essoussi. Feature selection using ge-
netic algorithm for big data. In The International Conference on Advanced Machine
Learning Technologies and Applications (AMLTA2018), pages 352–361, 2018.

[11] Saad Munir and M. Asif Naeem. Bil-fand: leveraging ensemble technique for effi-
cient bilingual fake news detection. International Journal of Machine Learning and
Cybernetics, Mar 2024.

[12] Ehtesham Hashmi, Sule Yildirim Yayilgan, Muhammad Mudassar Yamin, Subhan
Ali, and Mohamed Abomhara. Advancing fake news detection: Hybrid deep learning
with fasttext and explainable ai. IEEE Access, 12:44462–44480, 2024.

[13] Mahabuba Akhter, Syed Md. Minhaz Hossain, Rizma Sijana Nigar, Srabanti Paul,
Khaleque Md. Aashiq Kamal, Anik Sen, and Iqbal H. Sarker. Covid-19 fake news
detection using deep learning model. Annals of Data Science, Jan 2024.

[14] Chander Prabha, Meena Malik, Shalini Kumari, Neha Arya, Parul Parihar, and
Jaspreet Singh. Detection of fake news: A comparative analysis using machine learn-
ing. AIP Conference Proceedings, 3072(1):040014, 03 2024.

[15] Joy Gorai and Dilip Kumar Shaw. Semantic difference-based feature extraction tech-
nique for fake news detection. The Journal of Supercomputing, Jun 2024.

[16] Muhammed Baki Çakı and Muhammet Sinan Başarslan. Classification of fake news
using machine learning and deep learning. Journal of Artificial Intelligence and Data
Science, 4(1):22–32, 2024.

[17] Manvi Bohra, Indrajeet Kumar, and Kamred Udham Singh. Fake news detection
using different machine learning algorithms. In 2024 2nd International Conference
on Device Intelligence, Computing and Communication Technologies (DICCT), pages
45–50, 2024.

[18] Pooja Malhotra and S. K. Malik. Fake news detection using ensemble techniques.
Multimedia Tools and Applications, 83(14):42037–42062, Apr 2024.

[19] Kirandeep Kaur and Nirbhay Kashyap. Fake news detection using boosting ensemble
method. In 2024 14th International Conference on Cloud Computing, Data Science
Engineering (Confluence), pages 532–537, 2024.

[20] Mohammed A. Taha, Haider D. A. Jabar, and Widad K. Mohammed. Fake news
detection model basing on machine learning algorithms. Baghdad Science Journal,
21(1):35–45, January 2024. Published Online First: January 20, 2024. Received
07/03/2023, Revised 14/07/2023, Accepted 16/07/2023.

[21] Gregorius Airlangga. Comparative analysis of machine learning algorithms for detect-
ing fake news: Efficacy and accuracy in the modern information ecosystem. Journal of
Computer Networks, Architecture and High Performance Computing, 6(1):354–363,
Jan. 2024.

[22] Sheik Abdullah A, Parkavi R, P.Je Sai Kailash, and Deepthi Ramesh. Analysis of on-



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 38 of 39

line fake news using machine learning techniques. In 2024 Second International Con-
ference on Emerging Trends in Information Technology and Engineering (ICETITE),
pages 1–5, 2024.

[23] Azka Kishwar and Adeel Zafar. Fake news detection on pakistani news using machine
learning and deep learning. Expert Systems with Applications, 211:118558, 2023.

[24] Minjung Park and Sangmi Chai. Constructing a user-centered fake news detection
model by using classification algorithms in machine learning techniques. IEEE Access,
11:71517–71527, 2023.

[25] Mutaz A. B. Al-Tarawneh, Omar Al-irr, Khaled S. Al-Maaitah, Hassan Kanj, and
Wael Hosny Fouad Aly. Enhancing fake news detection with word embedding: A
machine learning and deep learning approach. Computers, 13(9), 2024.

[26] Mutaz A. B. Al-Tarawneh, Omar Alirr, and Hassan Kanj. Performance evaluation of
machine learning-based cyber attack detection in electric vehicles charging stations.
International Journal of Advanced Computer Science and Applications, 16(3), 2025.

[27] Sudhansu R Lenka, Sukant Kishoro Bisoy, Rojalina Priyadarshini, and Mangal Sain.
Empirical analysis of ensemble learning for imbalanced credit scoring datasets: a sys-
tematic review. Wireless Communications and Mobile Computing, 2022(1):6584352,
2022.

[28] Adel Mellit and Soteris Kalogirou. Assessment of machine learning and ensemble
methods for fault diagnosis of photovoltaic systems. Renewable Energy, 184:1074–
1090, 2022.

[29] Mutaz A. B. Al-Tarawneh, Khaled S. Al-Maaitah, and Ashraf Alkhresheh. Adaptive
ensemble selection for personalized cardiovascular disease prediction using cluster-
ing and feature selection. International Journal of Advanced Computer Science and
Applications, 16(3), 2025.

[30] Richard A. Berk. Classification and Regression Trees (CART), pages 1–65. Springer
New York, New York, NY, 2008.

[31] Ludmila I Kuncheva. Combining pattern classifiers: methods and algorithms. John
Wiley & Sons, 2014.

[32] Bernhard Scholkopf. Support vector machines: a practical consequence of learning
theory. IEEE Intelligent systems, 13, 1998.

[33] Thi Kha Nguyen and Thi Phuong Trang Pham. Predicting bankruptcy using machine
learning algorithms. Tp ch́ı Khoa hc và Công ngh-i hc à Nng, pages 6–9, 2018.

[34] Thiago José Lucas, Inaê Soares De Figueiredo, Carlos Alexandre Carvalho Tojeiro,
Alex Marino G De Almeida, Rafa l Scherer, José Remo F Brega, João Paulo Papa, and
Kelton Augusto Pontara Da Costa. A comprehensive survey on ensemble learning-
based intrusion detection approaches in computer networks. IEEE Access, 2023.

[35] Pierre Geurts, Damien Ernst, and Louis Wehenkel. Extremely randomized trees.
Machine learning, 63:3–42, 2006.

[36] Antonio Criminisi et al. Regression forests for efficient anatomy detection and local-
ization in ct studies, sep. 20, 2010, medical computer visions. recognition techniques
and applications in medical imaging.

[37] Earum Mushtaq, Aneela Zameer, and Asifullah Khan. A two-stage stacked ensemble



M. Al-Tarawneh et al. / Eur. J. Pure Appl. Math, 18 (2) (2025), 6087 39 of 39

intrusion detection system using five base classifiers and mlp with optimal feature
selection. Microprocessors and Microsystems, 94:104660, 2022.

[38] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In
Proceedings of the 22nd acm sigkdd international conference on knowledge discovery
and data mining, pages 785–794, 2016.

[39] Hamoud Aljamaan and Amal Alazba. Software defect prediction using tree-based
ensembles. In Proceedings of the 16th ACM international conference on predictive
models and data analytics in software engineering, pages 1–10, 2020.

[40] Mimusa Azim Mim, Nazia Majadi, and Peal Mazumder. A soft voting ensemble
learning approach for credit card fraud detection. Heliyon, 10(3), 2024.

[41] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth
IEEE International Conference on Data Mining, pages 413–422, 2008.

[42] Antonella Mensi and Manuele Bicego. Enhanced anomaly scores for isolation forests.
Pattern Recognition, 120:108115, 2021.