Early Multi-Class Ensemble Based Fake News Detection Using Content Features

doi:10.21203/rs.3.rs-2217954/v1

Download PDF

Research Article

Early Multi-Class Ensemble Based Fake News Detection Using Content Features

https://doi.org/10.21203/rs.3.rs-2217954/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 31 Dec, 2022

Read the published version in Social Network Analysis and Mining →

You are reading this latest preprint version

Nowadays, social media play an essential role in spreading the news with low cost and high speed in publishing, and easy availability. As anybody can publish some news, many fake materials are also published in this media. Multi-class detection of fake news (more than just false or true) have gained more attention, as some news are half-true. In addition, early detection of fake news has an important role, before their impacts on the society become great. This paper investigates an early detection of fake news using multi-class classification. This is achieved by extracting the content features from the news, such as sentiment and semantic features. The proposed model employs five classifiers (Random Forest, Support Vector Machine, Decision Tree, LightGBM, and XGBoost) as primary classifiers. Furthermore, the AdaBoost is used for the meta-learning algorithm to develop a stacking generalization model concerning fake news detection. Stacking generalization is an ensemble learning method that uses all data produced by the first-level algorithms. We trained our model with Politifact data in the evaluation, and the model performance was evaluated by their Accuracy, Precision, Recall, and F1 score. Performance improvements are visible over previous models, in both binary and multi-class classifications.

natural language processing

machine learning

fake news detection

ensemble learning

Today, social media has become an integral part of human life. The role of traditional information channels such as television and newspapers in receiving the news has diminished compared to the past. In this area, social networks such as Twitter and Facebook are the most popular. Social media lacks control over the trustworthiness of published information. This becomes important when we realize the direct impact of fake information on our lives.

People spread the news consciously or unconsciously without awareness of its validation. Due to a captivating title, the news may be published several thousand times without being validated. For example, the fake news that Barack Obama was injured in an explosion caused the value of US stocks to fall by $ 130 billion ¹.

Abusers can spread the fake news to get political and economic advantages, such as publishing that news to destroy an electoral rival. Therefore, reliable detection methods should be developed to prevent these evil intents, which can help people not fall into the trap of believing fake news. One of the most prominent examples of how fake news affects humans is USA presidential election in 2016. In the election, most USA residents were exposed to fake articles about Trump more than Clinton, but it was not an adequate reason to influence the election outcome. Due to the analysis of the poll result, it has been shown that the Republican voters were generally more inclined to the fake news articles ².

The main features of the news are represented in news content and sometimes have only these features at the early stage of the propagation. Therefore, extracting valuable content features helps identify the fake news lonely. All inventions and technologies are developing, and fake news also follows this rule. The fake news written style is continuously evolving, as are the methods of identifying the fake news. Detecting the fake news becomes problematic when the written style of fake news is similar to the real news. When we check out fact-checking websites, they assign different labels to the news. Traditionally, the news was collected in datasets with two true and false labels. But today, fake news writers use a combination of true and false paragraphs to make fake news more credible to the reader. This is a reason to provide models that can categorize fake news with multi-class labels. The fake paragraph is intentionally written for malicious purposes and is the most critical part of news paragraphs that mislead the reader. The data used to build the fake news detection model is collected from sources that have labeled the news differently. In previous works, the news was used only in binary for classification (the news with different labels, sometimes by specifying the threshold or removing other labels, has been used only in binary).

We used news collected from three reputable fact-checking websites and built a multi-class model to identify fake news. Therefore, we do not miss any information from the news that has labels different from true and false.

The main contributions of this paper are summarized below:

Using the most useful linguistic features like textual, sentiment, semantic, and readability features
Employing a new dataset that contains multi-source and multi-class fake news on a variety of topics
The proposed model uses a Stacking ensemble network with five main classifiers (Random Forest, SVM, Decision Tree, LGBM, and XGBoost)
Experiments were performed in two ways, binary and multi-class, then compared with each related literature.

The remainder of this paper is organized as follows. In section 2, the related literature is reviewed. Then the classifier algorithms are briefly introduced in section 3. In section 4, the ensemble learning methods is presented. How the model is developed is elaborated in section 5. Finally, the experimental results are reported, and the conclusion is drawn in sections 6 and 7, respectively.

Linguistic features are extracted from news content text directly and find essential information about fake news. There are many valuable Natural Language Processing (NLP) tools to analyze the collected information that can help us to fake news detection. For instance, Stanford CoreNLP^{^[2]} and NLTK^{^[3]} are NLP tools to extract knowledge from text. Vicario et al. ³ worked on polarization and bias between social media users who spread fake news online. In this paper, they tried to early polarize content detection and extracted textual features like semantic features and sentiment features. The number of characters, words, and sentences belongs to semantic features. They used linear regression, logistic regression, support vector machine, and K-nearest neighbor for classification.

Zhou et al. ⁴ investigated a theory-driven model of fake news detection, which proposed four different levels for content features. Characters per word, Sentences per paragraph, word per sentence, and Number of sentences used as Quantity features. Diversity features had considered, such as the percentage of unique words and verbs, and sentiment features like the percentage of positive comments. These are some content features that important for fake news detection. They provided a complete list of semantic features involved in their study. Their experiments have been conducted on PolitiFact and BuzzFeed datasets with binary labels. Classifiers used in their paper are SVM, RF, and XGBoost.

Message-based features are one of the features that were considered to detect fake news by Castillo et al. ⁵. It has been pointed out that message-based features can be Twitter-dependent or independent of Twitter. Twitter-independent features such as message length, question mark, and the number of emotional words are positive or negative. The Twitter-related quality mentions the hashtag and retweets the message. Unique sentiment features used in this paper are positive and negative word count in terms of sentiment and total sentiment score.

Rashkin et al. ⁶ compared the language of real news with humor, hoaxing, and advertising to obtain unreliable text features. They reached the language of real news with humor and hoaxing to obtain unreliable text features. In this study, Politifact data with six labels were used. Politifact statements often have a degree of honesty between true and false. It has been pointed out that false information follows a slight difference rather than an obvious structure. They used NLTK tools for lexicon analysis and reported that words exaggerate and superlative adjectives and modal verbs are used more in writing fake news. Content features have valuable and significant information to identify fake news written style. Finally, they reported macro averaged F1 score for development set on Politifact 2-class and 6-class data. The best performance was LSTM without LIWC features for 2-class data and Maximum Entropy (MaxEnt) for 6-class data.

Huang et al. ⁷ used a deep learning model to identify fake news. The proposed model is ensemble learning networks with four training models. Models are N-gram CNN, LIWC CNN, depth LSTM, and embedding LSTM and optimizing ensemble learning weights with Self-Adaptive Harmony Search (SAHS) techniques, as shown in Fig. 1. This SAHS algorithm is used to reach higher accuracy in the fake news detection model. After preprocessing step, they extract Uni-gram and Bi-gram for N-gram CNN. Besides, LIWC was used to get eigenvector of news for LIWC CNN, Grammar Parser produced sentence depths for depth LSTM, and embedding LSTM used tokenizing words. Three datasets (Buzzfeed, Satire, and Politifact) are used in this work.

Agarwal et al. ⁸ implemented an ensemble learning approach to merge various classification models with SVM, Naïve Bayes, k-nearest neighbor, CNN, and LSTM. They said combing these models makes fake news detection more accurate. NB, KNN, and LSTM classifiers were trained on two datasets (Liar and collected dataset from Kaggle) and tried to produce credibility scores in short news. The Lair dataset has six different labels, and this paper sets a threshold to change labels to binary form, then uses the binary dataset to make its model.

When investigating the related works, it turns out that they did not use up-to-date datasets. Almost all classifications were formed in binary, which can affect the efficiency of the built models. Fake news written form is progressing over time. Therefore, new models are needed to identify fake news outside of the binary format.

[2] https://stanfordnlp.github.io/CoreNLP/

[3] https://www.nltk.org

In this section, the basis of the algorithms used in this paper is described briefly.

3.1. Random Forest

A Random Forest (RF) algorithm fits several classifier trees on each dataset sample. For classification problems, the final class is selected by most trees. The number of trees built in the forest is controlled by "n-estimators" parameters ⁹.

3.2. SVM

Support Vector Machine (SVM) is a supervised machine learning algorithm that can be used for classification. The SVM makes one or set of hyper-planes in N-dimensional, and the planes classify the training data points of any class. The formula present in the following poses the optimization issue solved by SVM ¹⁰.

$$\underset{\omega ,b,\zeta }{\text{min}}\frac{1}{2}{w}^{T}w+C\sum _{i=0}^{n}\zeta$$

$${\gamma }_{i }\left({w}^{T}\varphi \left({x}_{i}\right)+b\right)\ge 1-{\zeta }_{i}$$

Where w^Tw represents the average vector, C is a regularization parameter, ${\zeta }_{i}$ represents the distance to the correct margin, $\varphi \left({x}_{i}\right)$ represents the transformed input space vector, b is a bias parameter, and ${\gamma }_{i }$ represents the i-th target.

The most popular kernel functions are the linear, polynomial, Radial Basis Function (RBF), and sigmoid functions ¹⁰. We use SVM algorithms with RBF kernel that function is:

$$k\left({x}_{i},{x}_{j}\right)= \text{e}\text{x}\text{p}(-\gamma {‖{x}_{i}-{x}_{j}‖}^{2})$$

3.3. Decision Tree

The Decision Tree (DT) makes classification models in the form of a tree structure. DT is a supervised learning method that predicts the target class by learning rules derived from the data features. One of the essential parameters in the DT algorithm is called "criterion." These parameters used to calculate the quality of a split and supported criteria are "Entropy" and "Gini" ¹¹. The following equation where give the Entropy ${p}_{i}$ is the probability that a point is in the subset of the dataset:

$$Entropy\left(P\right)= -\sum _{i=1}^{n}{p}_{i}{\text{log}}_{2}\left({p}_{i}\right)$$

And Gini index is given by the below equation:

$$Gini\left(P\right)=1-\sum _{i=1}^{n}{{(p}_{i})}^{2}$$

3.4. LightGBM

LightGBM is a gradient boosting algorithm that uses trees for classification. When there are many features in data, they can be used LGBM. It has several advantages like:

Handling large dataset
Efficient memory usage
Faster training speed
Higher accuracy

To solve the problem of a large number of features, uses gradient-based side sampling (GOSS) or Exclusive Features bundling (EFB) ¹².

3.5. XGBoost

XGBoost is one of the valuable and efficient distributed gradients boosting models. It is an end-to-end tree-boosting and supervised model. XGBoost prepares a parallel tree boosting that solves many classification problems accurately and fast ¹³.

Ensemble learning is a machine learning technique that combines outputs of basic learner algorithms and produces one optimal prediction model ¹⁴. Ensemble methods can be used to increase the outperformance of predictive models. The three typical ensemble learning methods are Stacking, Bagging, and Boosting.

Stacking comprises several different training models on the same data and uses another model to achieve the best combination of the prediction ¹⁵, as shown in Fig. 2 (a).

Bagging perhaps is the simplest method in ensemble learning with good efficiency. It uses different dataset samples and then combines them by taking a simple majority vote of the prediction ¹⁵, as shown in Fig. 2 (b).

Boosting trains models sequentially, where those methods mentioned above are trained in parallel. A new model is created and solving its incompetence in the previous step ¹⁵, as shown in Fig. 2 (c).

We create our model for early fake news detection in 3 main steps. Firstly, all the news in the dataset is preprocessed to make data ready for machine learning algorithms. Secondly, features are extracted from the title and the text of the news, for example, generated the new title for each news, and calculated similarities, sentiment scores, and quantity features are added to the list of content features that help identify fake news. Finally, a stacking ensemble network is built with five basic classifiers. Table 1 shows all components used in our model.

Table 1

Complete features list is extracted from the news
Features name	Category	Features name	Category
Similarities between the title and the text of the news	Semantic	Ann semantic score	Sentiment
Similarities between the topic modeling title and the text of the news	Semantic	Vader semantic score	Sentiment
Number of characters	Quantity	Positive words count	Sentiment
Number of words	Quantity	Negative words count	Sentiment
Number of sentences	Quantity	Coleman-Liau Index (CLI)	Readability
Number of capital words	Quantity	Gunning fog index (GFI)	Readability
Number of punctuations	Quantity	Automated Readability Index (ARI)	Readability
Avg. number of characters per word	Quantity	Flesch-Kincaid Grade Level (FKGL)	Readability
Avg. number of words per sentence	Quantity	Flesch Reading Ease Index (FREI)	Readability

5.1. Data preprocessing

The raw data in the dataset is prepared for machine learning algorithms by removing worthless pieces. At the first step, all news (text and title of news) converts to lower case, then extra spaces are replaced with one space, and some special characters like Â €, ™ are removed. At last, all English stop words were released by the NLTK library.

A typical algorithm used to fit textual content into machine algorithms for prediction is TF-IDF. The news was converted to a numeric vector with TF-IDF and N-grams in our work. Vectors are made with 1-gram, 2-gram, 3-gram, and 4-gram.

5.2. Feature generation

Content-based features are one of the first and most essential features that can be used to detect fake news. So, the features that can produce from the news text include 1- Generate new title for the news (Topic modeling), 2- Calculate similarity, 3- Calculate sentiment score, 4- Quantity textual features.

5.2.1. Topic modeling

The primary purpose of the news title is to absorb the reader's attention. When people see the attractive title encouraged to read the whole news article, and also news title is used to affect the reader's discernment of fake news ¹⁶. To solve the problem of mismatching the news title with the news text to mislead the reader, we use Topic modeling with NLTK and Gensim libraries. We can find discussed topics in the news text and then compare them with the original news titles with topic modeling. We use Wordnet of NLTK to understand the word's meaning and lemmatize the terms to get the root. Using Latent Dirichlet Allocation (LDA), find ten topics in the news, then find similarities between the original title and issues generated in the calculating similarity section with the topic modeling method ¹⁷.

It can be said that when the similarity between the original title of the news and the topics extracted from the text of the news is low, this news is prone to be fake because the author tried to create a wrong attitude in the reader by changing the title of the news. This is a novel feature that extracts from news content.

5.2.2. Calculate similarity

This part divides into two steps. Computing similarity between the original title and news text at the first step and the second step is similarity calculation between topic modeling output with the original news title. These two similarity values are added to the extracted features list.

5.2.3. Calculate sentiment score

We use two popular approaches for sentiment score calculation. Afinn ¹⁸ is a word list approach for analyzing the sentiment of the text. The similarity score range produced by Afinn is from − 526 to 282. Valence Aware Dictionary and sEntiment Reasoner (VADER) ¹⁹ is a rule-based tool for sentiment scores. This tool presented a negative sentiment score, a positive sentiment score, a neutral sentiment score, and a combined score for each news. The combined score was calculated by adding the three scores together and then normalizing between + 1 and − 1. Also, the number of positive and negative words (from the perspective of sentiment) is counted and added to the list of features.

5.2.4. Quantity textual features

Fake news written style can be captured by collecting quantity features. These features include the number of characters, the number of words, and the number of sentences. If we want to show how much the news is complex can use the average of characters per word and words per sentence. The best features in the news text can be extracted when the goal is early detection. Therefore, we prepared a complete list of these features shown in Table 1 and used them to perform better represent fake news to machine learning algorithms ⁴.

5.3. Building model

After preprocessing, vectors as output are prepared for classification, but there is another step. In feature extraction, we collected meaningful features which must be added to the prepared vectors. After doing this, those are ready to be used for machine training. Figure 3 has two stages; stage 1 shows preprocessing and features generation steps. Stage 2 represents an ensemble-based network with five basic classifiers (Random Forest, SVM, Decision Tree, LGBM, and XGBoost) that uses AdaBoost as a meta classifier. Figure 3 shows how to build a model from start to end.

6.1. Datasets

One of the most frequently used datasets in previous works is the Liar dataset ²⁰. Liar is a multi-class dataset with six different labels collected from traditional news outlets like TV or radio and campaigns. There are 12836 short statements with 'True,' 'Mostly-True,' 'Half-True,' 'Barely-True,' 'False,' and 'Pants of Fire' labels. The other valuable columns include statement, subjects, speaker, and speaker's job.

There are several fake news datasets like Buzzfeed corpus ²¹, Satire dataset ²², CREDBANK ²³, and FEVER ²⁴ that each of which has its characteristics, and researchers select all or part of them to use in their research. Still, one of the most striking features is that the news label is in binary format like True or False. For this reason, lots of papers built a binary fake news detection model. Multi-class datasets have been converted to binary datasets by thresholds and then used.

A new dataset was recently introduced that collected 24517 news from three fact-checking websites (Politifact.com, Snopes.com, and truthorfiction.com) from September 1995 to January 2021. Most of the collected news is related to Politifact with about 60 percent of the dataset, then Snopes with 35 percent standing in second place, and Truthofiction with just about 5 percent is in last place. These websites have different methods to set the label for the news. There are five standard labels in this dataset (True, Mostly-True, Half-True, Mostly-False, and False) for each news. This dataset distinguishes itself from others with two features: multi-class and up-to-date news ²⁵.

We selected this dataset because of collecting new data from reliable websites. The fake news writing style is progressing, and this development impacts the detection, so the more recent data can help make an efficient detection model. The dataset consists of 23935 (each row with a null value is removed), as shown in Fig. 4, and has five labels (True, Mostly-true, Half-true, Mostly-False, and False).

Our paper performs the experiments on Intel I7-4710HQ CPU, NVIDIA GeForce GTX 850M GPU, and 12 GB memory. One of the famous python libraries for classification, regression, and clustering is Scikit-learn which we use to create first-level classifiers. First-level classifiers that we use to create an ensemble network are Random Forest, SVM, Decision Tree, LGBM, and XGBoost. The Scikit-learn library also creates the stacking ensemble network. The stacking network comprises the output of individual classifiers and uses a meta classifier to compute the final prediction.

6.2. Evaluation

Different metrics were provided for evaluation, such as Accuracy, Precision, Recall, and F1-Score, so we compare models with them.

Accuracy: Defined as the number of correctly predicted data instances over the total number of cases.

$$Accuracy=\frac{\left|TP\right|+\left|TN\right|}{\left|TP\right|+\left|TN\right|+\left|FP\right|+\left|FN\right|}$$

Precision: Defined as the proportion of correctly predicted positive instances to the total positive ones.

$$Precision=\frac{\left|TP\right|}{\left|TP\right|+\left|Fp\right|}$$

Recall: Defined as the proportion of correctly predicted positive instances to all instances in the actual class.

$$Recall=\frac{\left|TP\right|}{\left|TP\right|+\left|FN\right|}$$

F1 Score: The weighted average of Precision and Recall.

$$F1-\text{S}\text{c}\text{o}\text{r}\text{e}=2\text{*}\frac{precision\text{*}recall}{precision+recall}$$

Besides the most popular evaluation metrics mentioned above, Cost Per Example (CPE) is one of the metrics used to show how much multi-class models cost for a wrong prediction. A cost matrix is constructed to evaluate a multi-class fake news classification model, as shown in Table 2. In the multi-class classification, this matrix is better displayed. Eq. (5) shows how to calculate the CPE ²⁶.

$$CPE=\frac{1}{N}\sum _{i=1}^{m}\sum _{j=1}^{m}CM\left(i,j\right)*C(i,j)$$

Where C(i,j) and CM(i,j) are confusion and cost matrices, N denotes the total number of test samples, and m represents the number of classes in the classification operation. A confusion matrix is a square matrix in which each column is assigned to an actual class, and each row is assigned to a predicted class. Each element in row i and column j, CM(i, j), indicated the number of not correctly classified samples that belonged to class i and were ranked in class j. The elements on the primary diameter of the matrix indicate the number of samples that are correctly classified. The cost matrix is structurally similar to the confusion matrix. Its values are set between zero and one, except that the C(i, j) element in this matrix is the penalty cost for incorrectly classifying an instance. Therefore, the principal diameters of matrix C always have a zero value because the principal diameter indicates the correct classification of the samples. The cost matrix is designed innovatively between 0 and 1, costing zero for the best case and costing one for the worst-case scenario. This matrix may not be the best, but it is a benchmark for evaluating our model. When the calculated penalty is closer to zero, the model receives a minor penalty and performs better ²⁶.

Table 2

Cost per example
	True	Mostly true	Half true	Mostly False	False
True	0	$\frac{1}{4}$	$\frac{1}{2}$	$\frac{3}{4}$	1
Mostly true	$\frac{1}{4}$	0	$\frac{1}{4}$	$\frac{1}{2}$	$\frac{3}{4}$
Half true	$\frac{1}{2}$	$\frac{1}{4}$	0	$\frac{1}{4}$	$\frac{1}{2}$
Mostly false	$\frac{3}{4}$	$\frac{1}{2}$	$\frac{1}{4}$	0	$\frac{1}{4}$
False	1	$\frac{3}{4}$	$\frac{1}{2}$	$\frac{1}{4}$	0

As shown in Table 3, the accuracy of five basic classifiers is reported separately. While we describe our model with the Accuracy, Precision, Recall, and F1-score, some works only reported part of these evaluation metrics, so a comparison has been made with the reported metrics.

Table 3

Accuracy of each model
Model	Accuracy
SVM	72%
Decision Tree	78%
Random Forest	79%
XGBoost	82%
LGBM	84%

We train our model with a multi-class dataset in two different ways. For evaluation with binary models built in Huang et al. ⁷, Shu et al. [23], Zhou et al. [4], and Palani et al. ²⁷ need to change the multi-class dataset to a binary dataset. Firstly, we convert labels of the dataset to binary form, which changes the mostly-true label to true and mostly-false to false and leave out the half-true labels then train the model with them. evaluation report is shown in Table 4. Secondly, we train the model with multi-class data and compare it with Rashkin et al. ⁶, who provided only F1-Score for comparison. Still, all evaluation metrics of our model are reported in Table 5. Each previous work performed its model with different datasets, but Politifact data is common within all of them. Therefore, in one part of our experiments, we train our model on Politifact data and then evaluate it. Train and test sizes are 80% and 20%, respectively.

Table 4

Comparison of the result of our model on Politifact data (binary classification)
	Politifact
	Accuracy	Precision	Recall	F1-score
Huang et al. ⁷	76	75	75	75
Shu et al. ²⁸	87	86	89	88
Zhou et al. ⁴	89	87	90	89
Palani et al. ²⁷	93	92	91	92
Stacking Ensemble Network	96.24	96.67	96.74	96.71

Table 5

Comparison of the result of our model on Politifact data (multi-class classification)
Methods	Politifact
	Accuracy	Precision	Recall	F1-score
Rashkin et al. ⁶	---	---	---	22
Stacking Ensemble Network	94.40	94.31	94.02	94.15

Our primary intent is to build a multi-class fake news detection model with diverse Politifact, Snopes, and Truthorfiction. These are the three most famous fact-checking websites ²⁵. After all experimental evaluations are reported, we describe the classification report on complete data in a multi-class form, as shown in Table 6. Also, the confusion matrix of this trained model brings in Table 7.

Table 6

Comparison of the result of our model
	All Data (Politifact + Snopes + Truthorfiction)
	Accuracy	Precision	Recall	F1-score
The Proposed Model Multi-Class Form	83.60	83.97	81.94	82.81
The Proposed Model Binary Form	91.52	93.21	93.65	93.43

Table 7

Confusion matrix result of our model on multi-class data
		True	Mostly True	Half True	Mostly False	False	CPE
Actual label	True	1641	112	38	53	4	0.0487
	Mostly True	112	600	41	23	3
	Half True	91	68	660	18	3
	Mostly False	71	51	28	522	1
	False	15	34	24	6	568
		Prediction label

This paper introduces a fake news model that can predict the news with two labels (true and false) and five labels (true, mostly-true, half-true, mostly-false, and false). This model helps us to understand how much news is reliable. Combining the best text classification algorithms with ensemble learning networks is valuable because it can achieve a more accurate model than each model alone. Today, one of the difficulties in detecting fake news is that their writing style is more similar to the real news. The evaluation results show that the ensemble learning outputs of multi-class prediction on all training data and Politifact data are 83% and 94% as F1-score, respectively. Also, our binary form outperforms 96% as precision. The primary threat of our work is the lack of context features that can be used in combination with components extracted from news content to train the model.

Ethical Approval and Consent to participate

This research received no specific funding from funding organizations in the public, commercial or non-profit sectors. Also, all the authors have participated in this research work with full satisfaction and all its parts have been reported with complete accuracy and transparency.

Consent for publication

This research work can be published with the consent of all its authors.

Human and Animal Ethics

Not Applicable

Availability of supporting data

Everyone can see our supporting data in this link: https://github.com/Sajjad75/Early-Multi-Class-Ensemble-Based-Fake-News-Detection-Using-Content-Features

Funding and Competing interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Authors' contributions

The first author was involved in designing the initial idea, designing the implementation method and building the model and programming parts of the classification and ensemble learning algorithms.

The second and third authors have collaborated in ideation, correction of the article and supervision of the work process.

The fourth author was involved in correcting model programming errors and providing optimal solutions.

Acknowledgments

Not Applicable

Rapoza K. Can âĂŸfake newsâĂŹ impact the stock market. 2017.
Allcott H, Gentzkow M. Social media and fake news in the 2016 election. Journal of economic perspectives. 2017;31(2):211-236.
Vicario MD, Quattrociocchi W, Scala A, Zollo F. Polarization and fake news: Early warning of potential misinformation targets. ACM Transactions on the Web (TWEB). 2019;13(2):1-22.
Zhou X, Jain A, Phoha VV, Zafarani R. Fake news early detection: A theory-driven model. Digital Threats: Research and Practice. 2020;1(2):1-25.
Castillo C, Mendoza M, Poblete B. Information credibility on twitter. Paper presented at: Proceedings of the 20th international conference on World wide web2011.
Rashkin H, Choi E, Jang JY, Volkova S, Choi Y. Truth of varying shades: Analyzing language in fake news and political fact-checking. Paper presented at: Proceedings of the 2017 conference on empirical methods in natural language processing2017.
Huang Y-F, Chen P-H. Fake news detection using an ensemble learning model based on self-adaptive harmony search algorithms. Expert Systems with Applications. 2020;159:113584.
Agarwal A, Dixit A. Fake news detection: an ensemble learning approach. Paper presented at: 2020 4th International Conference on Intelligent Computing and Control Systems (ICICCS)2020.
Ho TK. Random decision forests. Paper presented at: Proceedings of 3rd international conference on document analysis and recognition1995.
Bishop CM, Nasrabadi NM. Pattern recognition and machine learning. Vol 4: Springer; 2006.
L. Breiman JF, R. Olshen, and C. Stone. Classification and Regression Trees. 1984; https://scikit-learn.org/stable/modules/tree.html#mathematical-formulation.
Ke G, Meng Q, Finley T, et al. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems. 2017;30.
Chen T, Guestrin C. Xgboost: A scalable tree boosting system. Paper presented at: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining2016.
Opitz D, Maclin R. Popular ensemble methods: An empirical study. Journal of artificial intelligence research. 1999;11:169-198.
Rokach L. Ensemble-based classifiers. Artificial intelligence review. 2010;33(1):1-39.
Zhou X, Zafarani R. Fake news: A survey of research, detection methods, and opportunities. arXiv preprint arXiv:181200315. 2018;2.
Li S. Topic Modelling in Python with NLTK and Gensim. 2018; https://towardsdatascience.com/topic-modelling-in-python-with-nltk-and-gensim-4ef03213cd21.
Nielsen FÅ. A new ANEW: Evaluation of a word list for sentiment analysis in microblogs. arXiv preprint arXiv:11032903. 2011.
Hutto C, Gilbert E. Vader: A parsimonious rule-based model for sentiment analysis of social media text. Paper presented at: Proceedings of the International AAAI Conference on Web and Social Media2014.
Wang WY. " liar, liar pants on fire": A new benchmark dataset for fake news detection. arXiv preprint arXiv:170500648. 2017.
Potthast M, Kiesel J, Reinartz K, Bevendorff J, Stein B. A stylometric inquiry into hyperpartisan and fake news. arXiv preprint arXiv:170205638. 2017.
Rubin VL, Conroy N, Chen Y, Cornwell S. Fake news or truth? using satirical cues to detect potentially misleading news. Paper presented at: Proceedings of the second workshop on computational approaches to deception detection2016.
Mitra T, Gilbert E. Credbank: A large-scale social media corpus with associated credibility annotations. Paper presented at: Ninth international AAAI conference on web and social media2015.
Thorne J, Vlachos A, Christodoulopoulos C, Mittal A. Fever: a large-scale dataset for fact extraction and verification. arXiv preprint arXiv:180305355. 2018.
Rezaei S, Kahani M, Behkamal B. The Process Of Multi-Class Fake News Dataset Generation. Paper presented at: 2021 11th International Conference on Computer Engineering and Knowledge (ICCKE)2021.
Toosi AN, Kahani M. A new approach to intrusion detection based on an evolutionary soft computing model using neuro-fuzzy classifiers. Computer communications. 2007;30(10):2201-2212.
Palani B, Elango S, Viswanathan K V. CB-Fake: A multimodal deep learning framework for automatic fake news detection using capsule neural network and BERT. Multimedia Tools and Applications. 2021:1-34.
Shu K, Wang S, Liu H. Beyond news contents: The role of social context for fake news detection. Paper presented at: Proceedings of the twelfth ACM international conference on web search and data mining2019.

No competing interests reported.

Download PDF

Journal Publication

published 31 Dec, 2022

Read the published version in Social Network Analysis and Mining →

Editorial decision: Major revision
01 Dec, 2022
Reviews received at journal
14 Nov, 2022
Reviewers agreed at journal
14 Nov, 2022
Reviewers agreed at journal
09 Nov, 2022
Reviewers invited by journal
08 Nov, 2022
Editor assigned by journal
08 Nov, 2022
Submission checks completed at journal
01 Nov, 2022
First submitted to journal
30 Oct, 2022

You are reading this latest preprint version

	True	Mostly true	Half true	Mostly False	False
True	0	\(\frac{1}{4}\)	\(\frac{1}{2}\)	\(\frac{3}{4}\)	1
Mostly true	\(\frac{1}{4}\)	0	\(\frac{1}{4}\)	\(\frac{1}{2}\)	\(\frac{3}{4}\)
Half true	\(\frac{1}{2}\)	\(\frac{1}{4}\)	0	\(\frac{1}{4}\)	\(\frac{1}{2}\)
Mostly false	\(\frac{3}{4}\)	\(\frac{1}{2}\)	\(\frac{1}{4}\)	0	\(\frac{1}{4}\)
False	1	\(\frac{3}{4}\)	\(\frac{1}{2}\)	\(\frac{1}{4}\)	0

Early Multi-Class Ensemble Based Fake News Detection Using Content Features

Status:

Journal Publication

Version 1

Abstract

Figures

1. Introduction

2. Related Works

3. Classifiers

3.1. Random Forest

3.2. SVM

3.3. Decision Tree

3.4. LightGBM

3.5. XGBoost

4. Ensemble Learning

5. Proposed Approach

5.1. Data preprocessing

5.2. Feature generation

5.2.1. Topic modeling

5.2.2. Calculate similarity

5.2.3. Calculate sentiment score

5.2.4. Quantity textual features

5.3. Building model

6. Experiments

6.1. Datasets

6.2. Evaluation

7. Conclusion

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1