International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index Evaluating Machine Learning Algorithms for Detecting Online Text-based Fake News Content Deni Kurnianto Nugroho1 .
Marwan Noor Fauzy2.
Kardilah Rohmat Hidayat3 Department of Information System.
Faculty of Computer Science Universitas Amikom Yogyakarta Yogyakarta.
Indonesia deni@amikom.
id1 , marwannoorfauzy@amikom.
id2, kardilah.
rh@amikom.
AbstractAiThe rapid spread of disinformation and fabricated news across online platforms poses a critical risk to informed public engagement and the foundations of democratic governance.
This study examines how well different machine learning techniques can classify fake news, using textual features extracted through the Term FrequencyAeInverse Document Frequency (TF-IDF) method.
The analysis includes five commonly used algorithms like Logistic Regression.
Support Vector Machine (SVM).
Naive Bayes.
Random Forest, and XGBoost.
A publicly accessible dataset containing annotated real and fake news articles served as the basis for training and testing these models.
Dataset underwent extensive preprocessing, including tokenization, stopword removal, and TF-IDF vectorization, resulting in a sparse high-dimensional matrix of 5068 documents and 39,978 features.
Performance evaluation was based on multiple metrics:
train/test accuracy, misclassification rate, false positives/negatives, cross-validation mean score, and execution time.
Results showed that SVM and Logistic Regression achieved the highest test accuracy .
61% and 92.
27%, respectivel.
and exhibited robust cross-validation scores, indicating strong generalization ability.
In contrast.
Naive Bayes produced faster results but suffered from a high false positive rate and lower accuracy .
77%).
Random Forest and XGBoost demonstrated good predictive power but showed signs of overfitting and moderate misclassification rates.
These findings suggest that SVM and Logistic Regression are well-suited for fake news detection in textual datasets using TF-IDF features.
While traditional models remain effective, future work may explore deep learning approaches and context-aware language models to enhance detection accuracy across more complex and multilingual datasets.
This study contributes to the ongoing efforts to combat misinformation through automated, scalable, and interpretable machine learning techniques.
Keywords : fake news, text classification, tf-idf, machine learning, supervised learning INTRODUCTION The development of information technology has brought significant changes in the way people access and disseminate One consequence of this ease of use is the increasing spread of false information or fake news, particularly through online media platforms and social media .
This phenomenon not only disrupts the democratic process, as seen in the 2016 US elections .
, but also has the potential to cause social unrest, misinformation on health issues, and division within society .
Manually identifying fake news has become increasingly impractical due to the massive scale of online information and the speed at which it spreads.
As a result.
Machine Learning (ML) techniques have gained traction as an effective solution to this issue .
These techniques allow systems to learn from data and detect patterns that help distinguish between authentic and fabricated news based on textual features .
ML-based methods have shown promising results across a range of tasks, particularly in the domains of natural language processing (NLP) and automated text classification .
outperforms others in all scenarios .
, emphasizing the importance of comparative evaluation to determine the most suitable model for specific use cases.
In the context of online news, other challenges include variations in writing styles, biases in news sources, and the presence of clickbait and disguised opinions that are difficult to distinguish from fact .
Therefore, an effective classification approach must be able to capture the linguistic and semantic nuances of news text.
This paper investigates how different machine learning classifiers perform in the task of identifying fake news articles, using textual features derived from a well-known benchmark dataset.
The analysis places emphasis on standard evaluation criteria including accuracy, precision, and recall while also incorporating assessments of model consistency through five-fold cross-validation and the time required for training and prediction.
Among the models tested.
Logistic Regression and Support Vector Machine (SVM) consistently demonstrate robust results, particularly with regard to accuracy and recall, and exhibit stable performance across validation folds.
Naive Several commonly adopted algorithms for fake news Bayes offers an appealing balance of speed and accuracy, detection include Support Vector Machine (SVM).
Random making it a practical option for real-time or resourceForest (RF).
Logistic Regression (LR), and Naive Bayes constrained scenarios.
In comparison, ensemble approaches (NB) .
The effectiveness of each algorithm often depends such as Random Forest and XGBoost achieve strong on factors such as the chosen text representation method predictive outcomes, albeit with longer processing times.
TF-IDF or word embedding.
, the nature of the These results illustrate the balance practitioners must strike features, and the size and quality of the dataset.
Prior studies between accuracy, interpretability, and computational cost have also highlighted that no single algorithm universally when developing fake news detection systems.
Journal IJCIS homepage - https://ijcis.
net/index.
php/ijcis/index Page 259 International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index Beyond technical challenges, fake news detection also has social and ethical dimensions that cannot be ignored.
The algorithm used must be able to handle data bias, as well as consider the implications of misclassification, such as risks to freedom of expression or the reputation of the party being misreported .
Therefore, the development of an accurate and responsible fake news detection system requires a multidisciplinary approach, encompassing technical, social, and policy aspects.
This research seeks to support this direction by providing a quantitative analysis of the performance of various machine learning algorithms in the context of text-based fake news detection, as a basis for the development of more reliable and ethical systems in the .
Logistic Regression Linear models are often used for binary classification because of their ability to handle high-dimensional data efficiently .
II.
RESEARCH METHODS
Random Forest It is a decision tree-based ensemble algorithm that builds multiple models and aggregates the results to improve .
This study seeks to construct an automated system for identifying fake news through the application of machine learning techniques.
The primary goal of the classification model is to effectively differentiate between authentic (REAL) and deceptive (FAKE) news content using the textual information provided.
The methodological framework encompasses several key stages: exploratory data analysis, text cleaning and preprocessing, transformation of text into numerical features, model training, performance assessment, and examination of influential features that drive the modelAos predictions.
Dataset Description The dataset employed in this study originates from a publicly available corpus comprising 6335 news articles with assigned labels.
The class distribution is nearly even, consisting of 3171 articles identified as genuine (REAL) and 3164 labeled as false (FAKE).
Each data point includes three primary components: the headline .
itle of the articl.
, the main body of text .
ull news conten.
, and the classification label indicating its authenticity.
On average, each article contains approximately 776 words, offering sufficient textual depth for comprehensive content analysis.
Support Vector Machine (SVM) Applied with a linear kernel to maximize the inter-class margin in the TF-IDF feature space.
SVM is known to be robust to high dimensions and is suitable for text .
Naive Bayes It is a simple and very efficient probabilistic model for text, because it assumes independence between features .
Extreme Gradient Boosting (XGBoos.
It is one of the modern boosting methods that iteratively optimizes the errors of the previous model.
XGBoost is known for its high performance in various data mining competitions .
Each model was trained on a TF-IDF representation of the corpus, with standard hyperparameter settings to ensure fairness in performance comparisons.
Data Preprocessing The goal of text preprocessing is to enhance the quality of textual data representation prior to model input.
This process typically involves several steps: transforming all text to lowercase, eliminating punctuation and non-letter characters, removing frequently occurring stopwords, applying stemming using the Porter algorithm, and performing word Once the text is cleaned and standardized, it is converted into a numerical vector using the TF-IDF TF-IDF accounts for both the frequency of a word within a single document and its occurrence across the entire dataset .
, enabling the classification model to identify patterns by assigning greater importance to more informative Classification Algorithm This study tested five machine learning algorithms for the purpose of fake news classification:
Journal IJCIS homepage - https://ijcis.
net/index.
php/ijcis/index Model Evaluation To evaluate the classification performance, multiple complementary metrics were utilized, including training and testing accuracy, misclassification count, false positives, false negatives, and misclassification rate.
Accuracy measures the proportion of correct predictions out of all predictions, giving a general sense of overall False positives (FP) indicate real news articles incorrectly classified as fake, which is critical to minimize to avoid mislabeling credible information.
False negatives (FN) represent fake news articles incorrectly classified as real, which is equally important to reduce as they pose a risk of misinformation spread.
Misclassification rate complements accuracy by showing the proportion of incorrect predictions.
These metrics collectively provide a more nuanced understanding of each modelAos strengths and weaknesses, especially in balancing the trade-off between FP and FN in fake news detection, where both types of errors carry significant societal consequences.
In addition to static accuracy values, 5-fold crossvalidation was conducted to measure the generalizability of each model across different data splits.
The mean crossvalidation score (CV Mea.
reflects average performance on unseen folds, while 2y standard deviation (CV 2yStdDe.
offers insight into result variability and model stability .
This approach ensures that the evaluation is not biased toward a single train-test split and better represents realPage 260 International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index world performance.
The misclassification analysis focuses on two critical types of errors: false positives .
isidentifying fake news as rea.
and false negatives .
ailing to detect fake new.
, which are essential in evaluating model reliability in real-world Execution time for training and testing each model was also recorded to assess computational efficiency, particularly relevant when deploying models at scale.
For models that operate on linear decision boundaries, such as Logistic Regression and SVM, further analysis was performed on learned feature weights to identify influential words contributing to the classification of FAKE and REAL This interpretability aspect supports transparency and model explainability in sensitive domains like misinformation detection.
Figure 1 shows the frequency distribution of the number of words in article titles.
The average title length is 5 words, with a maximum length of 53 words and a minimum of just one word.
The distribution shows a relatively normal pattern with a slight skew to the The majority of articles have titles between 7 and 14 The length of an article's title can be an early indicator in detecting fake news.
Titles that are too short or too long can indicate an attempt to grab readers' attention, a common characteristic of clickbait content, often associated with fake .
Distribution of Article Content Length i.
RESULT AND ANALYSIS
This section evaluates the performance of various machine learning approaches commonly used in text classification tasks.
The models examined include linear classifiers such as Logistic Regression and SVM, probabilistic methods like Naive Bayes (NB), as well as ensemble-based techniques including Random Forest (RF) and XGBoost (XGB).
These models are applied to the task of fake news detection using TF-IDF for textual feature The evaluation considers key performance metrics accuracy, precision, recall.
F1-score, and confusion matrix and employs 5-fold cross-validation to assess generalization capabilities.
Further analysis of runtime and misclassification trends is conducted to identify model For interpretable algorithms, feature importance is also analyzed to uncover the most influential terms contributing to classification outcomes.
Exploratory Data Analysis (EDA) Before training the classification model, an exploratory analysis of the data structure was conducted to understand the characteristics of the news content, particularly the number of words in the title and the article text.
This is crucial for gaining initial insight into potential pattern differences between genuine and fake news.
Figure 2.
Frequency distribution of word counts in news article text Figure 2 shows the word count distribution of the articles.
The average article length is approximately 776 words, with the shortest article having 0 words and the longest exceeding 20,000 words.
While there are extreme outliers, most articles are concentrated below 1,500 words.
This distribution indicates that the news content in the dataset is relatively varied in length.
Extreme article lengths can impact the model's performance in feature extraction and should be taken into account during preprocessing or outlier .
Comparison of Article Length by Label .
Article Title Length Distribution Figure 1.
Frequency distribution of word counts in news article title Journal IJCIS homepage - https://ijcis.
net/index.
php/ijcis/index Figure 3.
Density distribution of news article word length by Page 261 International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index Figure 3 displays the density distribution of article lengths by label (FAKE and REAL).
It can be seen that articles labeled FAKE have a peak distribution at a lower word count compared to REAL articles.
This indicates that fake news tends to be shorter and more to the point.
This difference is consistent with previous research findings that suggest fake news is often structured in a more concise and emotional style to increase its spread on social media.
Conversely, genuine news tends to be longer and includes detailed information and valid citations.
While article length is not the sole factor in determining news authenticity, it can be used as an additional feature in the classification process, especially when used in conjunction with other semantic and syntactic features.
This large feature size reflects the diversity of vocabulary in the news corpus and poses a challenge in model training due to the high risk of overfitting if not handled properly.
The evaluation outcomes are summarized in Table 1, presenting a comparative analysis based on commonly used classification metrics.
Accuracy measures the overall correctness of predictions, precision captures how many of the predicted fake news items were actually fake, and recall evaluates the proportion of true fake news items that were successfully identified.
All metrics were reported using macro-averaging to account for class imbalance and provide a balanced view of model performance.
Preprocessing Stages Before training a fake news detection model, the text data undergoes a series of preprocessing steps to ensure the quality and consistency of its representation.
This step is crucial to ensure the machine learning model can capture relevant patterns and avoid distractions from noisy or redundant information.
The preprocessing process begins by converting all text to lowercase.
This aims to equalize the representation of semantically similar but letter-different words, such as "Berita" and "berita.
" Next, the text is cleaned, removing special characters such as numbers, punctuation, symbols, and URLs that do not contribute to the core meaning of the news content.
The cleaned text is then separated into tokens, or individual word fragments, through a tokenization To reduce the data dimensionality and reduce interference from overly common words, stopwords are also removed, such as "yang," "dan," "dalam," and so on.
These stopwords do not provide significant information in the context of fake news classification because they occur frequently in almost all documents.
Although some NLP approaches also employ stemming or lemmatization to convert words to their base form, this implementation did not employ this process.
This is because the text corpus used was already sufficiently clean and representative, and retaining the original word forms was considered to preserve a richer semantic context.
The preprocessed textual data was transformed into numerical form using the TF-IDF method.
This approach assigns weights to words based on how often they occur within an individual document compared to their occurrence across the full collection of documents.
Terms that are common within a specific article but uncommon throughout the broader corpus are given higher importance, reflecting their relevance to that particular document.
The final result of this stage is a TF-IDF matrix with dimensions .
8, 39.
, meaning there are 5,068 news articles and 39,978 unique word features resulting from Journal IJCIS homepage - https://ijcis.
net/index.
php/ijcis/index Classification Model Evaluation After completing the training process using multiple classification algorithms, the modelsAo ability to distinguish between real and fake news articles was carefully evaluated.
The assessment involved key performance indicators such as accuracy, precision, recall, and F1-score, calculated on a test set containing 1,267 samples.
To further ensure the reliability and generalizability of the results, a 5-fold cross-validation approach was implemented.
The models were further evaluated through 5-fold crossvalidation to assess their generalization performance.
The average cross-validation accuracy and the corresponding A2 standard deviations (CV A 2 St.
are reported to reflect the stability and consistency of each model across different data Furthermore, the total runtime of each model training and evaluation was included to provide insight into computational efficiency.
This comprehensive evaluation framework helps in balancing predictive performance, reliability, and practical considerations, especially for critical tasks such as fake news detection.
In addition to classification metrics, the table includes cross-validation results to assess the consistency of model The CV Mean refers to the average accuracy obtained across five folds, while 2ySTD represents twice the standard deviation, capturing performance variability and model robustness.
A lower standard deviation implies more stable and reliable predictions across different data splits.
Furthermore, the table reports each modelAos training runtime, offering insights into computational efficiency, a critical factor for real-world deployment.
By presenting these multiple dimensions of evaluation.
Table 1 facilitates a wellrounded comparison of the models' predictive capabilities and practical considerations.
Complementing this.
Table 2 provides precision, recall, and F1-score for both FAKE and REAL classes, enabling a deeper understanding of claspecific strengths and weaknesses.
This dual perspective ensures that model selection accounts not only for overall performance but also for the ability to handle the asymmetric costs of misclassifications in fake news detection.
Page 262 International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index Table 1.
Benchmarking Classification Models for Fake News Identification Model Accuracy MissFalse False Miss(Train/Tes.
classified Positive Negative classification Rate 95/0.
SVM
98/0.
Naive Bayes 89/0.
0/0.
XGBoost 99/0.
Mean Score Table 2.
Precision.
Recall, and F1-Score for Each Model Model Precision Recall F1-score (Fake/Rea.
(Fake/Rea.
(Fake/Rea.
90/0.
95/0.
92/0.
SVM
93/0.
94/0.
94/0.
Naive Bayes 98/0.
71/0.
82/0.
91/0.
92/0.
91/0.
XGBoost 92/0.
93/0.
92/0.
Macro Avg Recall .
Logistic Regression (LR) The Logistic Regression model achieved a training accuracy of 95.
17% and a testing accuracy of 92.
demonstrating strong generalization capabilities.
misclassified 98 out of 1,268 test samples, corresponding to a misclassification rate of 7.
The false positive count was 58, and false negatives were 40, indicating a relatively balanced misclassification pattern.
From a cross-validation perspective.
Logistic Regression had a mean CV score of 77%, with a 2y standard deviation of 0.
83%, suggesting stable performance across folds.
It also had a fast execution time of 2.
16 seconds, making it a practical choice for largescale applications.
Logistic Regression achieved balanced performance between precision and recall for both FAKE and REAL classes, with macro averages of 0.
92 across all three metrics.
The recall for FAKE news .
was slightly higher than precision .
, indicating the model was particularly effective in capturing fake news articles, albeit at the cost of a slightly higher false positive rate.
The F1-scores for both classes were identical .
, highlighting consistent predictive ability across categories.
This balanced profile makes Logistic Regression a robust choice where both classes are equally important.
Support Vector Machine (SVM) SVM delivered the highest test accuracy of all models, at 61%, and an impressive training accuracy of 98.
had the lowest misclassification count .
, equivalent to a misclassification rate of 6.
In terms of misclassification types, it produced 61 false positives and 20 false negatives, showing high sensitivity in detecting fake The model also had a high CV mean score of 93.
although its 2y standard deviation .
06%) was slightly higher than that of Logistic Regression, indicating more variability across folds.
However, the trade-off was computational efficiencyAithe execution time was the Journal IJCIS homepage - https://ijcis.
net/index.
php/ijcis/index Macro Avg Precision StdDev Execution Time .
Macro Avg F1score highest, at 208.
9 seconds, which may be a limitation in realtime settings.
SVM delivered the strongest overall macro averages .
for precision, recall, and F1-score, confirming its superior ability to distinguish between real and fake news.
Precision and recall values were well-balanced for both FAKE and REAL classes, with only minimal variance between them.
The high precision for REAL news .
suggests low false positive rates, while the high recall for FAKE news .
indicates strong sensitivity in detecting This combination of high accuracy and balanced class performance explains why SVM emerged as the top model in overall evaluation, despite higher computational cost.
Naive Bayes The Naive Bayes classifier exhibited the lowest overall performance, with a training accuracy of 88.
63% and a testing accuracy of just 84.
It misclassified 193 articles, translating to a misclassification rate of 15.
23%, the highest among all models.
More concerning was the very high number of false positives .
, suggesting that the model frequently misclassified fake news as real.
Although it had a very short execution time of 0.
05 seconds, its CV mean score was only 83.
43%, with a 2y standard deviation of 1.
reinforcing its lower reliability.
This model may only be suitable for preliminary baselines or low-resource scenarios.
Naive Bayes presented an imbalanced performance While it achieved an exceptionally high precision for FAKE news .
, recall for the same class was notably low .
, meaning the model often missed fake news instances.
Conversely.
REAL news classification showed the opposite trendAivery high recall .
but comparatively lower precision .
, indicating frequent false positives when labeling articles as REAL.
This imbalance is reflected in the lower macro averages .
recision: 0.
88, recall: 0.
F1:
While its near-instant execution time is appealing, this Page 263 International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index trade-off in class performance reduces its suitability for highstakes detection tasks.
Random Forest (RF) Random Forest achieved a perfect training accuracy of 100%, indicating potential overfitting.
Its testing accuracy 32%, slightly lower than that of SVM and Logistic Regression.
The model misclassified 110 test samples, yielding a misclassification rate of 8.
68%, with 70 false positives and 40 false negatives.
The CV mean score was 15%, accompanied by a 2y standard deviation of 1.
indicating reasonably stable performance.
The execution time was 9.
71 seconds, placing it in the mid-range among tested models.
While accurate, its tendency to overfit warrants careful tuning.
Random Forest demonstrated consistent and symmetric precision and recall scores for both FAKE .
91/0.
and REAL .
92/0.
The macro averages .
suggest stable performance without strong bias toward either Its balanced profile means it rarely sacrifices one classAos accuracy for the other, making it reliable for varied fake news detection contexts.
However, as observed earlier, the perfect training accuracy suggests some overfitting, meaning these results should be interpreted cautiously for unseen data.
XGBoost XGBoost also performed strongly, with a training accuracy of 99.
9% and a testing accuracy of 92.
misclassified 101 articles, with a misclassification rate of The model had 67 false positives and 34 false negatives, favoring real news detection slightly.
It showed robust generalization with a CV mean score of 91.
44% and 2y standard deviation of 1.
However, it was relatively slower than Logistic Regression and Random Forest, 64 seconds to complete.
XGBoost thus provides a good balance between accuracy and generalization but at a higher computational cost.
XGBoost maintained balanced metrics across classes, with macro averages of 0.
92 for precision, recall, and F1score.
It achieved slightly higher recall for FAKE news .
than for REAL news .
, indicating a mild emphasis on detecting misinformation.
The marginally higher precision for REAL news .
compared to FAKE .
also suggests a slight leaning toward reducing false positives for legitimate content.
These nuanced strengths make XGBoost appealing for applications prioritizing misinformation detection while minimizing false alarms.
recall, reducing its robustness for high-stakes fake news Random Forest and XGBoost performed reasonably well with strong class-wise metrics, but Random Forest showed signs of overfitting and XGBoost was slower than simpler alternatives.
These results highlight that model selection should consider not only accuracy, but also class-wise precision, recall, and F1-score alongside generalization ability and computational efficiency, especially in real-world news classification systems where the impact of false positives and false negatives can differ significantly.
VI.
CONCLUSION
This study investigates the capability of several machine learning algorithms in detecting fake news, using textual features derived from the TF-IDF method.
The models analyzed represent different categories of learning techniques, including linear models such as Logistic Regression and SVM, a probabilistic approach like Naive Bayes, and ensemble-based methods such as Random Forest and XGBoost.
A structured workflow was adopted throughout the study, encompassing exploratory data analysis, text preprocessing, model construction, and performance evaluation.
This approach enabled a detailed comparison of the strengths and weaknesses of each classifier in addressing binary classification problems involving real versus fake news.
The experimental results show that SVM and Logistic Regression deliver the most consistent performance, achieving test accuracies of 93.
61% and 92.
respectively, with relatively low misclassification rates.
Notably.
SVM demonstrated the highest model stability, as evidenced by the highest average cross-validation score (CV Mean Scor.
07%, although at the cost of significantly longer computation time.
In contrast.
Naive Bayes, while computationally efficient, produced lower accuracy .
77%) and a higher tendency for false positives .
misclassifying fake news as rea.
, which poses risks in misinformation detection.
Random Forest and XGBoost also performed strongly, with accuracies above 91%.
However, they exhibited slightly higher misclassification rates compared to Logistic Regression and SVM.
Random Forest achieved 100% accuracy on the training set, indicating potential overfitting, whereas XGBoost offered a balanced trade-off between accuracy, validation stability, and computational efficiency.
Overall, the findings suggest that SVM and Logistic .
Result Summary Overall.
SVM provided the best generalization Regression are the most reliable choices for fake news performance in terms of test accuracy and misclassification detection tasks based on the dataset used, considering rate, while also maintaining balanced precision and recall accuracy, misclassification types, and cross-validation across both FAKE and REAL classes, though at a high stability.
The TF-IDF approach effectively captured essential computational cost.
Logistic Regression offered a balanced textual features, resulting in a high-dimensional feature trade-off between accuracy, generalization, efficiency, and matrix .
8 y 39,.
that proved suitable for classification.
consistent performance for both classes.
Naive Bayes, while extremely fast, showed an imbalance between precision and Journal IJCIS homepage - https://ijcis.
net/index.
php/ijcis/index Page 260 International Journal of Computer and Information System (IJCIS) Peer Reviewed Ae International Journal Vol : Vol.
Issue 03.
August 2025 e-ISSN : 2745-9659 https://ijcis.
net/index.
php/ijcis/index Subsequent studies may consider incorporating more sophisticated methods for textual representation, including word embeddings and transformer-based architectures like BERT.
Additionally, assessing model effectiveness on multilingual and heterogeneous datasets could enhance the modelsAo ability to generalize across broader contexts.
THANK-YOU NOTE
The authors would like to extend their sincere appreciation to Universitas Amikom Yogyakarta for the ongoing support, research infrastructure, and academic atmosphere that made this study possible.
We are especially thankful to the Department of Information Systems.
Faculty of Computer Science, for their technical guidance, encouragement, and constructive input throughout the We are also grateful to Kaggle for providing access to the Fake News Prediction Dataset, which served as a critical resource for the development and evaluation of our classification models in detecting misinformation.
In addition, we acknowledge the valuable insights and feedback from fellow researchers, academic colleagues, and anonymous reviewers, which greatly enhanced the clarity and rigor of this paper.
Lastly, we would like to recognize the global open-source and data science communities whose tools, frameworks, and shared knowledge played a vital role in enabling efficient data preprocessing, modeling, and result visualization, contributing significantly to the reproducibility and reliability of our research outcomes.
REFERENCES