Journal of Artificial Intelligence and Data Science (JAIDA) Vol. 4, No. 1, pp.

22-32, 2024
e-ISSN: 2791-8335 Research Article
https://dergipark.org.tr/pub/jaida Received: 17/04/2024 Accepted: 03/06/2024

Classification of Fake News Using Machine Learning and Deep Learning

Muhammed Baki ÇAKI 1, Muhammet Sinan BAŞARSLAN 2,*

Abstract

The rapid advancement of technology has led to an increase in the spread of fake news, which has a detrimental effect on
people in various fields, particularly in their daily lives. The negative impacts of fake news can be mitigated through the
use of artificial intelligence. The development of AI technologies has made the detection of fake news a prominent area
of research within natural language processing. This study explores style-based fake news detection using machine
learning and deep learning methods. The texts were processed using natural language processing techniques and
investigated with different models on the open-source ISOT dataset. The models utilised text processing, text
representations (TF-IDF, word2Vec), and different machine learning (ML) methods (K-Nearest Neighbor, Naïve Bayes,
Logistic Regression) as well as Long Short-Term Memory (LSTM). The performance of the models was evaluated using
accuracy (Acc), precision (P), recall (R), and F1-score. Among the tested models, the LSTM model demonstrated the
highest performance, with an accuracy of 99.2%. The development of state-of-the-art methods for text representation and
classification, including preprocessing in text classification, and the application of these methods in practical settings can
significantly reduce the prevalence of fake news.

Keywords: Deep learning, Fake news detection, Machine learning, Style based detection.

1. Introduction
Artificial Intelligence (AI) is a field that is divided into many sub-headings with its potential and is the subject
of many researches in order to produce better solutions to our problems. Natural Language Processing (NLP),
Machine Learning (ML) and Deep Learning (DL) are the main sub-topics of AI. Fake news is one of the problems
we want to solve. The rapid spread of false content produced for various reasons causes social and economic
damage to individuals, organizations and societies. This problem is growing with the increasing speed of
communication. Misinformation and disinformation have negative effects on society. Therefore, new and
effective methods are needed to detect and prevent fake news.

The main purpose of our study is to contribute to existing studies to find solutions to this problem with AI. In
order to classify and distinguish between fake and real news, linguistic features of news texts are processed and
analyzed with NLP. Then ML and DL models are built. After the models are trained, prediction is made for the
given news text to be real or fake. In this study, various models are built using different NLP techniques and ML
algorithms and the results obtained are analyzed. The results of the study show that NLP and ML models have a
significant potential in fake news detection.

In the second part of the study, similar studies in the literature are presented. The third section discusses the
dataset, preprocessing, vectorization, ML, DL and performance criteria. The fourth section describes the
experimental setup. Section five presents the experimental results. The discussion and conclusion in sections six
and seven provide an overall assessment and future works.

2. Related Works
Similar studies on fake news detection in the studied ISOT dataset will be described in this section. In their
study, the researchers created models with Logistic Regression (LR), Naïve Bayes (NB), Support Vector Machine
(SVM), Random Forest (RF) and deep neural network. They achieved 91% Acc with neural [1] . After GloVe,
the best performance with 92% Acc was obtained by using Linear Support Vector Machine (LSVM) as a classifier
[2]. After vectorization with Term Frequency - Inverse Document Frequency (TF-IDF) on ISOT dataset, classifier

*Corresponding author
Muhammed Baki ÇAKI; Istanbul Medeniyet University, Faculty of Engineering and Architecture, Computer Engineering Department, Türkiye; e-mail:
[email protected]; 0009-0005-2651-4047
Muhammet Sinan BAŞARSLAN; Istanbul Medeniyet University, Faculty of Engineering and Architecture, Computer Engineering Department, Türkiye;
e-mail: [email protected]; 0000-0002-7996-9169

22
 Çakı et al. / JAIDA vol (2024) 22-32

models were created for fake news detection with various ML algorithms. Among these models, the best result
was obtained with Decision Tree (DT) with 96.8% Acc [3]. They created models with ML methods such as SVM,
LSVM, K-Nearest Neighbor (KNN), DT on the data of fake news collected by ISOT and themselves. LSVM gave
the highest Acc result with 92% [4]. After vectorization with the Word2Vec method called Maithi-Net, they
obtained 97.28% Acc result in fake news detection with this method [5]. They obtained 74% Acc with NB after
CBOW [6]. Word2Vec obtained 82.67% Acc with conditional random fields (CRF) classifier after CBOW [7].
They obtained 99% Acc in a study on fake news classification with ensemble learning after TF-IDF [8]. If we
look at other fake news studies other than this dataset; They obtained 99.10% Acc in fake news classification
model by hybridizing Recurrent Neural Network (RNN) and LSTM on Liar dataset [9]. In their study on the
detection of false news in the pandemic, they obtained 96.19% Acc and 95% F1 with the Convolutional Neural
Network (CNN) they proposed by optimizing hyperparameters after embedding methods such as GloVe [10].

It is seen that ML methods are frequently used after TF-IDF, Word2Vec text representation methods on ISOT
dataset [2] and similar content data related to fake news. After the text representation methods TF-IDF and
Word2Vec, which are frequently studied in the literature, ML (KNN, NB, LR), and DL (LSTM) models were
created for the classification of fake news.

The contribution of this study to the detection of fake news on ISOT, an open source shared and balanced
dataset, is listed below:
 It is investigated which of the popularly preferred text representation methods such as TF-IDF and
Word2Vec has more impact on the performance of the models.
 The holdout discrimination results of the models built with classical ML (KNN, SVM, NB, LR) and
DL (LSTM) are investigated.

3. Materials and Methods

In this section, dataset, preprocessing, vectorization, ML, DL, and performance criteria will be explained. In
this study, a text analysis-based approach is adopted for automatic fake news detection. In this approach, the
characteristics of the texts are extracted and processed by NLP methods, then modelled and predicted by ML
methods. The field of study is in the combination of ML, DL, and NLP fields in Figure 1.

Figure 1. Fields of AI

The experimental steps carried out in the study are given in Figure 2.

Figure 2. Study Pipeline

23
 Çakı et al. / JAIDA vol (2024) 22-32

As seen in Figure 2, the dataset was preprocessed and then subjected to text representation (TF-IDF,
Word2Vec). Then the model was created with LSTM, KNN, NB, LR with 75%-25% training-test separation. F1,
P, R, Acc were used to evaluate the models.

3.1. Dataset

In the study, ISOT Fake News Dataset [2], which consists of fake and real news data created by the researchers
with news collected from the internet between 2016-2017, was used.

The distribution of the dataset, which consists of 44,898 news in total, is given in Table 1. According to the
researchers, news with true content was collected from the reuters.com website, while news with false content
was collected from various websites marked as unsafe by Polifact [2]. Table 1 provides information about the
content of the dataset.
Table 1. Dataset
News Number of articles Subjects
Type Articles size
Real 21417 Government-News 1570
Politics-News 11272
Type Articles size
US News 783
Fake Left-news 4459
23481
Politics 6841
News 9050

Figure 3 shows a sample image from the dataset.

Figure 3. Summary Visualization of the Dataset

Figure 3 shows the title of the news item, the text of the news item, the date of publication of the news item
and the class label of whether it is fake or real. Figure 4 shows the graph of class distribution.

Figure 4. News Class Distribution

In Figure 4, according to the class distribution of the dataset; we see that the skewness coefficient is calculated
as -0.08 and the kurtosis coefficient is calculated as -1.18. Since the skewness is very close to zero, we can accept
the distribution as symmetric. In the light of this result, the dataset is balanced.

24
 Çakı et al. / JAIDA vol (2024) 22-32

3.2. Text Representation

In this section you will find information about text representation.

3.2.1. Term Frequency Inverse Document Frequency

The method called Term Frequency - Inverse Document Frequency is based on the principle of extracting the
attributes of the text by weighting each word in the text according to its importance. In this method, the importance
of words is determined by analyzing how many times they occur in the examined text and how many times they
occur in other texts. The TF-IDF representing the term in sentence t, document d is given in equation (1) [8].

𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡𝑖𝑚𝑒𝑠 𝑡𝑒𝑟𝑚 𝑡 𝑎𝑝𝑝𝑒𝑎𝑟𝑠 𝑖𝑛 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡 𝑑

𝑇𝐹(𝑡, 𝑑) = (1)
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡𝑒𝑟𝑚𝑠 𝑖𝑛 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡 𝑑

D is the collection of all documents (corpus), the addition of 1 to the denominator is to prevent the term from
dividing by zero if it is not found in any document IDF is given in equation (2) [8].

𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑐𝑜𝑟𝑝𝑢𝑠 𝑁

𝐼𝐷𝐹(𝑡, 𝑑) = 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑐𝑜𝑛𝑡𝑎𝑖𝑛𝑖𝑛𝑔 𝑡𝑒𝑟𝑚 𝑡+1
(2)

The TF-IDF in the document is given in Equation (3) [9].

𝑇𝐹 − 𝐼𝐷𝐹(𝑡, 𝑑, 𝐷) = 𝑇𝐹(𝑡, 𝑑) × 𝐼𝐷𝐹(𝑡, 𝐷) (3)

According to this equation, frequent occurrence of a word in the relevant document increases its importance.
If it is a common word in other documents, it decreases its importance. In this way, stopwords in documents also
become unimportant.

3.2.2. Word2Vec

Word2Vec is a method of converting words into vectors of real numbers using artificial neural networks.
Words with close meaning are also numerically close in vector representation. In this way, the semantic proximity
and context information of the words are kept [11].

3.3. Machine Learning

In this section, ML methods are described.

3.3.1. K-Nearest Neighbors

KNN algorithm is a lazy learning algorithm used in classification and regression problems in ML. In the space
where the data points are represented, prediction is made based on the distance of the relevant point to other
points. For the classification task, the distances to the k nearest points are calculated. It is predicted as belonging
to the class with the least total distance. The reason why it is categorized as a lazy learning algorithm is that there
is no learning phase before the data to be predicted arrives. It takes two basic parameters, 'number of neighbours'
and 'distance metric' [12].

The number of neighbours is the value 'k', which is also in the name of the algorithm. Distance is calculated
with the k nearest neighbours. The distance metric is the algorithm to be used to measure the distance. The most
commonly used equation for distance calculation is given in (Euclidean distance) equation (4).

𝑑(𝑥, 𝑦) = √∑𝑛𝑖=1(𝑥𝑖 − 𝑦𝑖 )2 (4)

25
 Çakı et al. / JAIDA vol (2024) 22-32

3.3.2. Naïve Bayes

NB Classifier is a probability-based prediction method that uses a simplified version of Bayes' Theorem in
probability. It is often used in classification tasks. Bayes' Theorem allows the calculation of the probability of
event A occurring when event B occurs; when the probabilities of event A occurring, event B occurring, event B
occurring when event A occurs are known. Its formula is given in equation (5) [13]:

𝑃(𝐴) ∗ 𝑃(𝐵|𝐴)
𝑃(𝐵) = 𝑃(𝐵)
(5)

In the NB Classifier, the denominator part of the Bayes equation is ignored since the aim is to find the class
with high probability instead of finding the exact value. For a two-class classification task, the probabilities of the
data belonging to classes X and Y are calculated with the help of the equation. Whichever class the probability of
belonging to is calculated to be higher, is predicted to belong to that class.

3.3.3. Logistic Regression

LR is an algorithm frequently used in classification problems in ML. It is based on probability-based class

prediction by fitting the data to the logistic function. It is more suitable for binary classification task. Multi-class
classification can also be performed. Figure 5 shows a visual of LR [14].

Figure 5. LR [14].

3.4. Deep Learning

In this section, LSTM, one of the DL methods used in this study, will be explained.

3.4.1. Long Short-Term Memory

Long short-term memory is a DL architecture that is an advanced version of the RNN model. In RNN, as each
output affects the next input, a memory structure is formed. This memory is short term. In long inputs, the effect
of past data on the new input decreases rapidly and disappears (gradient vanishing problem). In LSTM
architecture, input, output and forget gates are used in addition to RNN. In this way, by creating a short-term and
long-term memory structure at the same time, context information can be preserved in long inputs such as
paragraphs. LSTM architecture is frequently used in the development of sequence-based prediction systems such
as anomaly detection and time series [15]. Figure 6 shows the LSTM architecture.

26
 Çakı et al. / JAIDA vol (2024) 22-32

Figure 6. LSTM Architecture [15].

3.5. Performance Criteria

In this section, the metrics used in the evaluation of the performance criteria used for model evaluation and the
confusion matrix that is used to obtain this metrics are explained.

3.5.1. Confusion Matrix

The confusion matrix is constructed by comparing the predicted and actual values of the test data. In each cell,
the total number of samples belonging to that cell is recorded. In this way, the test result of the model can be
analysed on a single table [16].

 True Positive (TP) if the true value of the data is positive and predicted as positive,
 False Negative (FN) when the true value is positive and estimated as negative,
 False Positive (FP) when the true value is negative and estimated as positive,
 A negative true value and a negative predicted value constitute True Negative (TN) cases.

Figure 7 shows a visualization of the confusion matrix.

Figure 7. Confusion Matrix [17]

Various metrics are calculated to measure how well the created models make predictions. The predictions
made by the model are processed to the relevant part in the confusion matrix. Metrics are calculated using the
relevant fields in the confusion matrix. The four most important metrics used for evaluating classification tasks
are explained

27
 Çakı et al. / JAIDA vol (2024) 22-32

Accuracy:

It is the ratio of the model's correct predictions to all predictions [18-19]. This ratio is given in Equation (6).
𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐 = 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁 (6)

Precision:

It is the metric that measures the success ratio of the model when it predicts the outcome as positive [19]. This
metric is given in Equation (7).
𝑇𝑃
𝑃 = 𝑇𝑃+𝐹𝑃 (7)

Recall:

It is the metric that measures the extent to which the model can accurately detect situations that are actually
positive [20]. This metric is given in Equation (8).
𝑇𝑃
𝑅 = 𝑇𝑃+𝐹𝑁 (8)

F1:

In cases such as measuring high Acc and low P values in an unbalanced dataset, the Acc value may be
misleading about the success of the model. F1 metric is obtained by the harmonic mean of P and R values and
indicates the stability of the prediction [21]. This metric is given in Equation (9) [18].
𝑃∗𝑅
𝐹1 = 2 ∗ (9)
𝑃+𝑅

4. Experimental setup

The work was done in Google Colab [22], which allows the Python programming language [23] to run on a
Jupyter notebook. In this chapter, preprocessing of texts, TF-IDF and Word2Vec followed by modelling with ML
and DL will be explained.

4.1. Preprocessing

The operations performed within the scope of the study will be explained in this section. Attributes other than
text and label have been removed. Label attribute is a binary data type containing 0 for false news and 1 for true
news. The preprocessing processes are listed below:
1. The news source at the beginning of real news has been removed.
2. Letters and characters other than the '@' sign were removed from the news texts.
3. News texts were converted to lower case.
4. Stopwords in the news texts were removed.
5. The words in the news texts were stemmed.

4.2. Text Representation

Since the models to be created cannot process text data, the text must be matrixised and given as input to the
model. For this reason; TF-IDF with its implementation in sci-kit learn library, Word2Vec text representation
methods were used with its implementation in the gensim library. The Word2Vec parameters used in the study
are presented in Table 2.

28
 Çakı et al. / JAIDA vol (2024) 22-32

Table 2. TF-IDF and Word2Vec parameters

Parameters Value
Vector size 200
Window 5
Word2Vec Min count 5
Methods (sg) CBOW (0)
Workers 4

4.3. Model Creation

In the study, a total of seven models were created with TF-IDF and Word2Vec text representation method,
KNN, NB, LR, and LSTM. The data to be used to train and test these models are divided into 75% training and
25% test data. In the study, sklearn library was used to create models with KNN, LR, and NB, which are traditional
ML methods, and keras library was used for LSTM, which is a DL model.

4.3.1. Machine learning model

In the study, for the creation of ML models, Multinomial NB Classifier and LR Classifier implementations in
the sci-kit learn library were used to create models with default parameters. For the KNN model, the
implementation of the same library was used and the grid search algorithm was used for hyperparameter
optimization.

The TF-IDF vectorization method was tested with the grid search algorithm for the number of neighbours (k)
parameter of the generated KNN model for values (1-10) and the optimal value was observed to be 1. It was
observed that the optimum k value for the model created with Word2Vec vectorization method was 5. The distance
metric parameter was chosen as "euclidean".

4.3.2. Deep learning model

In the study, an LSTM model was created using the LSTM module in the Keras library. The model was
optimized by creating and comparing models with different hyperparameters. Used and preferred hyperparameters
are shown in table 3.

Table 3. TF-IDF and Word2Vec parameters

Output Dimension Neuron Dropout Epochs Batch size Loss Function Opitimizer Activation Function
100, 200 10-100 0.2, 0,3 5, 8, 10 32, 64 binary crossentropy adam sigmoid
*Optimum values in bold

Confusion matrix and score metrics obtained from confusion matrix were used to measure the model
performance. The loss functions and Acc values of the training and validation data were monitored to observe
that overfitting/underfitting situations do not occur in the training phase.

4. Experimental Results

Dataset is splitted as 75% and 25% for training and testing respectively. The results of models created after
text representation with TF-IDF and Word2Vec are given in Table 4 and Table 5.

Table 4. Results with TF-IDF

KNN NB LR
Acc 79.3 92.3 97.4
P 72.8 91.4 97.5
R 95.2 93.8 97.5
F1 82.5 92.6 97.5

As seen in Table 4, the model created with LR is ahead of the other models in Acc, P, R, and F1 in 75%-25%
hold-out separation after TF-IDF.

29
 Çakı et al. / JAIDA vol (2024) 22-32

Table 5. Results with Word2Vec

KNN NB LR LSTM
Acc 94.3 89.2 97.1 99.2
P 96.0 88.2 97.5 98.8
R 92.8 91.3 96.8 99.6
F1 94.4 89.7 97.2 99.2

As seen in Table 5, the model created with LSTM is ahead of the other models in A, P, R, and F in 75%-25%
hold-out discrimination after Word2Vec. When Table 4 and Table 5 are evaluated together, the results obtained
with Word2Vec in KNN are ahead. However, in LR and NB, the difference between TF-IDF and Word2Vec is
not much compared to KNN.

LSTM was the model that gave the best results among all models. The graphs of train-validation Acc and loss
values of the LSTM neural network are given in Figure 8.

Figure 8. Confusion Matrix

As predicted, the success rate of the LSTM model exceeded the classical ML algorithms. However, it was
observed that the training time of the model was considerably high compared to others.

5. Conclusion and Discussion

In the study where various traditional ML, DL models and text representation methods were compared for the
Fake News Detection task, it was observed that the best result was obtained with the DL Model, the performance
of the KNN model was more affected by the vectorization method, and NB and LR models obtained close and
good results in both vectorization methods. In particular, the LR Model was found to be the best model in terms
of efficiency for this binary text classification study. Similar studies on the same dataset are given in Table 6. Acc
results were used because the dataset is balanced.

30
 Çakı et al. / JAIDA vol (2024) 22-32

Table 6. Previous studies on the ISOT dataset

Authors Text Representation Model Acc (%)

[2] GloVe LSVM 92.0

[3] TF-IDF DT 96.8

[4] BOW LSVM 92.0

[5] Word2Vec CNN 97.28

[6] BOW NB 74.0

[7] Word2Vec CRF 82.6

[8] TF-IDF Ensemble learning (DT) 99.0

This study Wod2Vec LSTM 99.2

In the models created for the classification of fake news with ML (KNN, NB, LR) and DL (LSTM) after TF-
IDF and Word2Vec on the fake news dataset (ISOT), the best result was obtained with Word2Vec LSTM with
99.2% ACC. The model created in the study is a model that competes with the literature as seen in Table 6. The
models created with Word2Vec are more successful than the models created with TF-IDF in most cases. This
situation will be investigated in future studies by working on different datasets.

In the future, the process can be repeated with different data sets and the hyperparameters of the LSTM model
can be further optimized. In addition, newer and successful state-of-the-art models such as BERT, RoBERTa can
be used. As these models have been pre-trained with very large datasets, they have brought great advances in the
fields of NLP and ML. For this reason, it is predicted that the success of the study will increase.

Declaration of Interest
The authors declare that there is no conflict of interest.

Author Contributions
Muhammed Baki ÇAKI; data analysis, experiments and evaluations, manuscript draft preparation. Muhammet Sinan
BAŞARSLAN; defining the methodology, evaluations of the results, and original draft, supervision.

References
[1] C. K. Hiramath and G. C. Deshpande, “Fake News Detection Using Deep Learning Techniques,” in 2019
1st International Conference on Advances in Information Technology (ICAIT), IEEE, Jul. 2019, pp. 411–
415. doi: 10.1109/ICAIT47043.2019.8987258.
[2] H. Ahmed, I. Traore, and S. Saad, “Detection of Online Fake News Using N-Gram Analysis and Machine
Learning Techniques,” 2017, pp. 127–138. doi: 10.1007/978-3-319-69155-8_9.
[3] F. A. Ozbay and B. Alatas, “Fake news detection within online social media using supervised artificial
intelligence algorithms,” Physica A: Statistical Mechanics and its Applications, vol. 540, p. 123174, Feb.
2020, doi: 10.1016/j.physa.2019.123174.
[4] S. Minaee, N. Kalchbrenner, E. Cambria, N. Nikzad, M. Chenaghlu, and J. Gao, “Deep Learning--based
Text Classification,” ACM Comput Surv, vol. 54, no. 3, pp. 1–40, Apr. 2022, doi: 10.1145/3439726.
[5] D. Muduli, S. K. Sharma, D. Kumar, A. Singh, and S. K. Srivastav, “Maithi-Net: A Customized
Convolution Approach for Fake News Detection using Maithili Language,” in 2023 International
Conference on Computer, Electronics & Electrical Engineering & their Applications (IC2E3), IEEE, Jun.
2023, pp. 1–6. doi: 10.1109/IC2E357697.2023.10262664.
[6] M. Granik and V. Mesyura, “Fake news detection using naive Bayes classifier,” in 2017 IEEE First Ukraine
Conference on Electrical and Computer Engineering (UKRCON), IEEE, May 2017, pp. 900–903. doi:
10.1109/UKRCON.2017.8100379.

31
 Çakı et al. / JAIDA vol (2024) 22-32

[7] A. Priyadarshi and S. K. Saha, “Towards the first Maithili part of speech tagger: Resource creation and
system development,” Comput Speech Lang, vol. 62, p. 101054, Jul. 2020, doi: 10.1016/j.csl.2019.101054.
[8] I. Ahmad, M. Yousaf, S. Yousaf, and M. O. Ahmad, “Fake News Detection Using Machine Learning
Ensemble Methods,” Complexity, vol. 2020, pp. 1–11, Oct. 2020, doi: 10.1155/2020/8885861.
[9] A. K. Shalini, S. Saxena, and B. S. Kumar, “Automatic detection of fake news using recurrent neural
network—Long short-term memory,” Journal of Autonomous Intelligence, vol. 7, no. 3, Dec. 2023, doi:
10.32629/jai.v7i3.798.
[10] M. Akhter et al., “COVID-19 Fake News Detection using Deep Learning Model,” Annals of Data Science,
Jan. 2024, doi: 10.1007/s40745-023-00507-y.
[11] J. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, “Distributed representations of words and
phrases and their compositionality,” in Advances in Neural Information Processing Systems, 2013.
[Online]. Available: https://proceedings.neurips.cc/paper
[12] R. Ahmed, M. Bibi, and S. Syed, “Improving Heart Disease Prediction Accuracy Using a Hybrid Machine
Learning Approach: A Comparative study of SVM and KNN Algorithms,” International Journal of
Computations, Information and Manufacturing (IJCIM), vol. 3, no. 1, pp. 49–54, Jun. 2023, doi:
10.54489/ijcim.v3i1.223.
[13] T. Öztürk, Z. Turgut, G. Akgün, and C. Köse, “Machine learning-based intrusion detection for SCADA
systems in healthcare,” Network Modeling Analysis in Health Informatics and Bioinformatics, vol. 11, no.
1, p. 47, Dec. 2022, doi: 10.1007/s13721-022-00390-2.
[14] H. Canlı and S. Toklu, “Design and Implementation of a Prediction Approach Using Big Data and Deep
Learning Techniques for Parking Occupancy,” Arab J Sci Eng, vol. 47, no. 2, pp. 1955–1970, Feb. 2022,
doi: 10.1007/s13369-021-06125-1.
[15] R. Vankdothu, M. A. Hameed, and H. Fatima, “A Brain Tumor Identification and Classification Using
Deep Learning based on CNN-LSTM Method,” Computers and Electrical Engineering, vol. 101, p. 107960,
Jul. 2022, doi: 10.1016/j.compeleceng.2022.107960.
[16] H. Canli and S. Toklu, “Deep Learning-Based Mobile Application Design for Smart Parking,” IEEE
Access, vol. 9, pp. 61171–61183, 2021, doi: 10.1109/ACCESS.2021.3074887.
[17] M. Z. Khaliki and M. S. Başarslan, “Brain tumor detection from images and comparison with transfer
learning methods and 3-layer CNN,” Sci Rep, vol. 14, no. 1, p. 2664, Feb. 2024, doi: 10.1038/s41598-024-
52823-9.
[18] S. N. Başa and M. S. Basarslan, “Sentiment Analysis Using Machine Learning Techniques on IMDB
Dataset,” in 2023 7th International Symposium on Multidisciplinary Studies and Innovative Technologies
(ISMSIT), IEEE, Oct. 2023, pp. 1–5. doi: 10.1109/ISMSIT58785.2023.10304923.
[19] F. Kayaalp, M. S. Basarslan, and K. Polat, “TSCBAS: A Novel Correlation Based Attribute Selection
Method and Application on Telecommunications Churn Analysis,” in 2018 International Conference on
Artificial Intelligence and Data Processing (IDAP), IEEE, Sep. 2018, pp. 1–5. doi:
10.1109/IDAP.2018.8620935.
[20] Öztürk, T., Turgut, Z., Akgün, G. et al. Machine learning-based intrusion detection for SCADA systems in
healthcare. Netw Model Anal Health Inform Bioinforma 11, 47 (2022). https://doi.org/10.1007/s13721-
022-00390-2
[21] Ardaç, H.A., Erdoğmuş, P. Question answering system with text mining and deep networks. Evolving
Systems (2024). https://doi.org/10.1007/s12530-024-09592-7
[22] Google LLC, “Colab.” https://colab.research.google.com/. Accessed 1 Feb 2023
[23] Python, “Python.” https://www.python.org/downloads/. Accessed 1 Feb 2023

32