Detection of COVID-19 Fake News in Online Social Networks with the Developed CNN-LSTM Based Hybrid Model

In this study, which was carried out using the COVID-19 news dataset, fake news was detected. In the literature, studies using machine learning methods are in the majority. In this study, an LSTM-CNN based model was used. In the proposed hybrid model, 80% of the data in the data set was used for training the proposed hybrid model. 20% of the data randomly allocated in the dataset was evaluated during the testing process of the hybrid model. In addition, the matrices obtained after the data cleaning and glove process were classified using different machine learning methods. In addition, the performance of machine learning methods on the COVID-19 dataset was tested in the study. A rough flow diagram of the LSTM-CNN based model and other machine learning methods is given in Figure 1.

1.png

Figure 1. Block diagram of the model

2.1 Dataset and preprocessing

The COVID19-FNIR dataset used in the study was created by Saenz et al. [15]. The created dataset includes news about the COVID-19 pandemic. Researchers' data were collected from regions of India, the United States, and Europe. There are 3795 fake and 3793 true news in the dataset.

8 rows related to the data in the dataset are given in Table 1. These samples were randomly selected from the dataset.

Table 1. Text examples

In the study, first of all, the data in the data set was carried out from the preprocessing step. In this step, numbers, punctuation marks, Stopwords words, etc. steps have been carried out. The text preprocessing steps are shown in Figure 2.

2.png

Figure 2. Preprocessing steps

Artificial intelligence-based methods do not directly operate on the text. First of all, it is of great importance to prepare the data in the data set in a format that the machine can understand. Various preprocessing steps can be applied to the data before the data in the dataset is converted into a machine-understandable format. These preprocessing steps have a great impact on the performance of the models. In the study, punctuation marks and numbers in the dataset were cleaned first. Then, case transformations were applied to the data in the data set so that the models did not perceive the words containing upper and lower case letters as different words. Also, the datasets are cleared of too many repetitive stop-words (the, is, on, etc.). Since the performance at this stage can directly affect the performance of the models, this stage is of great importance in text processing studies.

Also, the word cloud method was used for word representation. The word cloud representation of the 500 most repeated words in the raw data set is given in Figure 3.

3.png

Figure 3. A word cloud representation of 500 words in the unprocessed dataset

The word cloud representation of the 500 most repetitive words of the preprocessed data set is given in Figure 4.

4.png

Figure 4. Word cloud representation of the 500 words in the processed dataset

2.2 CNN architectures, layers and LSTM networks

CNN networks are frequently used in studies such as signal processing, natural language processing, emotion recognition, segmentation, classification, and object recognition due to their high performance in computer vision. CNN architectures are one of the most used architectures recently. One of the biggest reasons why CNN architectures have been used very frequently in recent times is that the manually extracted feature extraction process is done automatically in CNN networks in classical methods. Since feature extraction in CNN networks is done automatically, there is no need for expert knowledge in these architectures [16]. Unlike classical methods, CNN architectures do the training process directly on the model. In addition, there is no need for a preprocessing step in CNN architectures.

Certain layers are used when creating CNN networks. These layers used in CNN networks work in a forward direction. There is no standard for the number and order of layers in the creation of CNN networks. The most commonly used layers in CNN networks are convolution, pooling, dropout, flatten and dense layers. In addition, activation functions are frequently used in CNN networks.

Convolution layer is the most basic layer of CNN architectures. In this layer, feature maps are obtained from the data by applying filters to the data. Since these filters are applied to the data from the beginning to the end of the input, feature maps are obtained from the data. With the filters applied to the input data, the depth of the meshes is increased and therefore it is aimed that the meshes give more accurate results. The output size obtained after the convolution process is calculated by Eq. (1) [17].

$O=\frac{(i-k)+2 p}{s}+1$            (1)

In Eq. (1), O stands for output size, i is input size, k is filter size, s is the number of steps, and p is padding.

The convolution operation is calculated by Eq. (2).

$O(i, j)=(D * K)(i, j)=\sum_{m} \sum_{n} D(m, n) K(i-m, j-n)$      (2)

In Eq. (2), O represents the output after the convolution operation, the K kernel (kernel), the D input data, and the "*" symbol represents the convolution operation.

The pooling layer, on the other hand, is used in CNN networks to reduce the size of the feature maps obtained in the convolution layers and to reduce the diversity. With the pooling layer, computational costs and the input size of the next layer are reduced. The main purpose of this layer is to get better feature maps [18]. The size of the feature map obtained after the pooling layer is calculated by Eqns. (3)-(5).

$G_{2}=\frac{G_{1}-F}{S}+1$      (3)

$Y_{2}=\frac{Y_{1}-F}{S}+1$       (4)

$D_{2}=D_{1}$     (5)

G₂ is the width of the new data size, G₁ is the width of the input image, Y₂ is the height of the new data size, Y₁ is the height of the input image, D₂ is the depth of the new image, D₁is the depth of the input image, F Filter size, and S is the number of steps.

In the literature, maximum pooling and average pooling are mostly preferred for the pooling process.

The dropout layer is used to minimize the memorization situation in the networks. This memorization situation that occurs during the training of the networks is also called overfitting. With the dropout layer, it is aimed that some neurons have less information about each other. Thanks to the dropout layer, neurons are less affected by each other's weight changes [19].

The flatten layer, on the other hand, depends on all areas of the previous layer. The Flatten layer takes the feature maps as input and makes this feature maps suitable for the classification process. Multidimensional feature maps are converted to a one-dimensional matrix in this layer [20].

The backpropagation algorithm is used to learn historical data in RNN networks. However, when backward calculations are made in RNN networks, vanishing gradient and exploding gradient problems may arise. Also, in RNN architectures, the next string is difficult to predict due to context gaps between strings. Hochhreiter and Schmidhuber developed LSTM networks to eliminate the disadvantages of RNN networks in their work. The most important difference that distinguishes LSTM networks from RNN networks is that LSTM networks have cells that hold previous state and input information. Thanks to these cells in LSTM networks, it is determined which data will be kept and which data will be deleted [21, 22].

In this study, a hybrid model is proposed by using CNN and LSTM networks together.

2.3 Shallow models and glove technique

The amount of data kept in databases has been increasing rapidly in recent years. The processing of this data held in datasets is of great importance. Unprocessed raw data has no value. The processing of data in datasets is very important for both the public and private sectors. institutions or companies can make inferences from these data and make plans according to the results they have obtained. Processing and analyzing the raw data held in data sets is quite difficult with traditional methods. Therefore, with the development of technology, artificial intelligence-based models are used more. Shallow models are frequently used in the processing and analysis of raw data held in datasets. These models are accepted in the literature and used in many studies.

In this study, 7 different Shallow models were used to detect fake news. Shallow models are designed in 2 stages. The first stage is the stage where the models are trained, and the second stage is the stage where these models are tested. 80% of the data in the data set was used to train the Shallow models. The models were also tested with the remaining 20% dataset in the dataset. The performance metrics obtained in these models were compared with each other and with the proposed CNN-LSTM based model.

The first of the models used in the study is k-nearest neighbors (KNN). KNN is a method that aims to predict the class based on the information in which class the nearest neighbors of the vector formed by the independent variables of the value to be estimated are in. is a method [23].

The Random Forest method, on the other hand, is based on tree structures. It is frequently used in both classification and regression problems [24].

In the Gradient Boosting (GB) and Adaboost methods, inferences are made from the collection of trees obtained as a result of giving different weights to the data set. Initially, all observation values are the same. As the tree community begins to grow, the weights are rearranged. The weight of misclassified observations is often increased. Because of this, trees gain the ability to self-regulate [25, 26].

The logistic Regression method is frequently used in classification and regression problems. This model is frequently used in the classification of categorical and numerical data [27].

Another model used in the study is the Discriminant Analysis model. In this method, in order to distinguish the classes from each other, the distribution of the classes is examined and the difference between their mean values is used [28].

A CNN-LSTM based model is proposed in the study for fake news detection. Different Shallow models were used in order to compare the performance values obtained in the proposed model. These models can be widely used in many fields such as image processing, audio processing, signal processing, text processing.

Text processing applications, the next step after the data preprocessing step is the word representation step. How to represent the text is one of the critical steps in text processing studies. Since the texts are not directly understood by the machines, in this step the texts are converted into numerical expressions that the machines can understand. Therefore, word representation methods are used to convert the text into numerical expressions. There are different word representation methods used in the literature. In this study, the glove method was used for word representation.

The working logic of unsupervised algorithms is to use the statistical information of the data. However, despite using this statistical information in matrix decomposition methods, semantic relationships cannot be captured. There is no semantic step in the logic of such models. The glove method is a method developed to eliminate these disadvantages. The Glove method, developed by Pennington et al., is a word representation method that takes its name from the initials of the words "Global Vectors for Word Representation". The glove is a word representation method that is trained on global word counts, thus allowing statistics to be used more effectively [29].

2.4 Developed CNN-LSTM based hybrid model

The COVID-19 epidemic has had a great impact on people since 2019. In this process, besides the health of people, their psychology has also deteriorated. The effect of fake news spread on social media on the deterioration of people's psychology is quite high. The widespread use of online social media platforms increases the number of fake news. It is very important that this news can be detected using computer-aided systems. At study, it is aimed to detect COVID-19 fake news. In the study, a hybrid model based on CNN-LSTM was developed. In addition, the performance values of different machine learning methods were measured and compared with the proposed model. The proposed model is detailed in Figure 5.

5.png

Figure 5. CNN-LSTM based hybrid model

In the hybrid model developed to detect fake COVID-19 news, 1 Embedding, 2 Convolution, 2 Pooling, one Dropout, and LSTM layers were used. RELU is preferred as the activation function. Parameter values in the proposed model are presented in Table 2.

Table 2. Parameters of the CNN-LSTM based hybrid model

The COVID-19 epidemic has been affecting all over the world since 2019. This epidemic affects both people's health and psychology. Especially in recent years, with the widespread use of social media platforms by users, the number of fake news made on these platforms is increasing day by day. It is very important to detect fake news made on the Internet using computer aided systems. Although machine learning methods are frequently used in text processing problems, CNN and LSTM structures have been used in recent years. In this study, a CNN-LSTM based hybrid model is proposed for the detection of fake COVID-19 news. When the performance metrics obtained in the proposed model are examined, the values obtained show that the proposed model can be used in fake news detection and text classification problems.