Ensemble-Based Machine Learning Approach for Detecting Arabic Fake News on Twitter

Fake news spreads differently on social media. Traditional News starts with well-established media businesses with fact-checking and editorial standards. Unlike social media, where anybody can publish, these institutions are limited, making it easier to regulate and manage news accuracy. Millions of users create diverse content, making management and accuracy difficult. Additionally, speed and reach in the second difference. Traditional news takes longer to propagate since it’s edited and validated. Fake news spreads quickly on social media. The rapid spread of viral disinformation makes it impossible to control. Control and verification are the third difference. Traditional media are responsible for accuracy. Social media users must verify without a central process. The fourth difference is user involvement and accessibility. Traditional news consumption is passive, with few outlets. User interaction is possible on social media. Creating, sharing, and commenting on content makes the environment more interactive. NLP and ML are needed to recognize social media fake news automatically. Due to controlled environments, traditional news sources need less automated systems [1]. For fake news, word disambiguation, sentiment analysis, and spam detection, NLP and ML classify and organize text. These technologies are essential because of social media’s global reach and enormous text data. Traditional media is controlled and accountable, but social media is user-driven. This change greatly affects fake news identification and management [2-5]. Due to social media’s broad use, fake news detection (FND) is crucial in the digital era. An enlarged view of fake news’ financial, political, and social effects: Financial markets can be impacted by fake news. Irrational investor behavior can cause market collapses due to misinformation about a company, economic policy, or market conditions. False information affects customer behavior, harming businesses and economies. Fake product or service news can damage a company’s reputation or increase its value. Economic effects: Fake news can damage financial institutions and markets, causing long-term economic damage. Political Effects: Fake news may swing elections by influencing voter behavior. Polarization of Society: False political information can deepen social differences and foster mistrust and enmity amongst political groupings. Fake news weakens democratic societies by undermining public trust in democratic institutions and procedures. Public Health: Misinformation about COVID-19, immunization, and medical therapies can harm public health. Social Harmony: Fake news can cause hatred, violence, and misunderstanding between communities, causing conflict. Media and Information Credibility: Fake news undermines reputable news sources and the public’s ability to tell the truth from fiction [6, 7]. These detrimental effects can be mitigated by effectively detecting bogus news. FND methods detect and stop misleading information before it harms [8]. Due to its complexity, this work requires technology solutions and collaboration between governments, social media platforms, news organizations, and the public. The goal is to develop an informed and discriminating online community that can reject misinformation. Research and advancement in fake news identification are crucial. Fake news affects financial markets, among other things [9], Rumors can destabilize markets. Foolish reactions to fake news are becoming more widespread as consumers decide based on knowledge [10, 11].

Countering erroneous information has led to the development of fact-checking websites and computational methods. However, these remedies have limitations. Methods of Computing: These technologies identify fake news by analyzing article wording. They may look for writing style, phrase building, and word choice patterns associated with misinformation. Contextual Limitations: Computational methods may misinterpret data context, resulting in inaccuracies. Misinforms’ Change: Fake news producers may change their methods to escape detection by these technologies. Implementation Complexity: Developing effective algorithms for altering data is challenging and resource-intensive [12]. Verification websites like PolitiFact and Snopes: These websites manually verify claims, particularly those viral on social media or in the media. Rich in Resources: Fact-checking is laborious and requires skilled human fact-checkers, making it hard to keep up with the huge amount of content published daily. Scope Limit: While some fact-checking groups focus on politics, they may ignore other fields with frequent erroneous information. Fact-checking takes time; thus incorrect information can spread before being challenged. Public Perception and Faith: In politically sensitive situations, fact-checkers’ prejudice may affect public faith. Some solutions combine human and machine fact-checking. These have issues too: Integrating automated and human verification processes efficiently is tough. Balancing Scalability and Accuracy: Automation is scalable, but precision and dependability are challenging to guarantee. In conclusion, fact-checking websites and computational detection methods are useful tools in fighting fake news, but they are not perfect. They struggle with scope, intricacy, domain limits, and public opinion. This highlights the need for continual innovation and progress in false news detection methods and larger public education and media literacy efforts [13].

NLP and ML are employed in fake news identification because they can process and evaluate massive amounts of data, including unstructured Web data. This article discusses these methods, their shortcomings, and how new ways can enhance them: Common NLP/ML Fake News Detection Methods: NLP systems detect false news language in article text. This may include reviewing writing style, word choice, and sentence structure. Analysis of sentiment the emotional tone of an item can indicate bias or deception [14]. The distribution of linguistic elements in fake and authentic news can show misleading patterns [15]. It analyzes how fake news spreads across networks and assesses it as true or false. It examines textual features and theoretical replies to determine deceit. Most research focuses on detecting and classifying fake news on Facebook and Twitter [16]. Text and metadata information are used to train ML models to categorize news as real or false. Modern NLP and ML Techniques’ Limitations: Automated systems can misidentify subtle fake news due to context and nuance issues. Fake news spreaders can adapt to avoid detection by these methods. While some models perform well in their trained area, they may not generalize well. The quality and diversity of training data determine ML model performance. Biased or inadequate training data can affect model performance. Improvements New Methods: Advanced NLP techniques may help understand news context and intricacy. Creating models that work across domains and fake news kinds. Including Network Dynamics: Consider how information moves across networks to verify news beyond textual analysis. Using NLP and ML with human verification or other data sources to improve accuracy. Developing systems that can adapt to misleading methods.

In conclusion, NLP and ML have improved fake news identification, although they have limits. These include difficulties with nuanced and contextually rich data, domain generalization, and the need for large, unbiased training datasets [12, 17, 18]. New methods improve analysis and include more data and methods to address these issues. This methodological growth underscores the dynamic and difficult nature of fighting fake news online.

The following are the important contributions made in this work:

This work presents an efficient model for detecting Arabic fake news content using multiple base classifiers, utilizing an ensemble to aggregate outputs for final prediction.

Because of the platform’s complexity, noise, bias, and dynamic nature, traditional Twitter false news detection tools, such as rule-based systems and machine learning techniques, may suffer.

The study addresses the growing incidence of fake news on social media, focusing on Arabic. It provides an ensemble-based machine-learning strategy for detecting bogus news on Twitter, incorporating multiple classifiers trained on diverse variables. This approach improves the performance and robustness of fake news detection. With this arrangement, the paper is now: Section 2 explores Related works, while Section 3 describes the proposed method used in developing the classification model. Section 4 describes the results and discussion of the proposed method, and Section 5 concludes the study.

The following learning algorithms are part of the primary block diagram for the proposed approach. Figure 1 depicts the proposed ensemble method and our suggested technique for assessing the effectiveness of false news detection classifiers.

1.png

Figure 1. Proposed ensemble method

3.1 Data collection

Compile a dataset of Arabic tweets categorized as either true or fraudulent news. Use various sources to ensure the dataset is broad and covers various topics.

3.2 Preprocessing

In this paper, the proposed method for preprocessing the dataset is done in several phases to ensure that the original text, which includes noise and numerous undesired features, was cleaned in multiple stages, as shown in Table 1:

1- D0, the original dataset

2- D1, the first clean data, which comprises Remove Stop Words, Remove Noise, Normalize, and Punctuations, will improve the accuracy of the proposed approach to discern between true and fake news.

3- D2, the new data that resulted from the second cleaning stage by applying Root steaming techniques to reduce the data set size for significantly faster learning.

4- D3, the final cleaning stage known as light steaming, is a little reduction for the dimensionality of the data set to speed up the procedure. The suggested model will include all the above data sets and compare the results to determine the best.

Table 1. Multi-stage preprocessing of Twitter news from a data set

3.3 Feature extraction

This study uses machine learning to convert textual expressions into vector expressions for mathematical operations in Natural Language Processing (NLP) studies. It extracts word-to-vector and Skip-Gram features from preprocessed data using semantically superior Word2Vec. One well-liked NLP feature extraction technique is Word2Vec. Vector word embedding is taught to Word2Vec models. Vectors represent word associations and semantic meanings. Large text datasets are used to train Word2Vec models. Words are predicted by the algorithms based on their context or the other way around. Word2Vec makes use of Skip-Gram and CBOW architectures. CBOW: Uses context words to predict the target word. Skip-Gram anticipates context terms for target words. After training, a high-dimensional vector (of several hundred dimensions) is assigned to each corpus word. In this space, words with similar meanings have close vectors. Word2Vec is trained to vectorize any text. These vectors can be used in sentiment analysis, topic modeling, and other NLP applications; Word2Vec is the recommended technique. Word2Vec is superior to bag-of-words models for meaning-based tasks because it captures semantic relationships between words. In contrast to one-hot encoding, Word2Vec represents densely, reducing computing overhead. Unlike previous methods, Word2Vec can capture word meanings in various contexts. Word2Vec models learned on one dataset can be applied to another for transfer learning. Word embedding from Word2Vec can be tailored to suit NLP workloads.

3.4 Model for ensemble learning

A machine learning technique known as ensemble learning teaches weak learners to work together to solve problems for better outcomes. There are three types: Bagging, Boosting, and stacking. Bagging focuses on producing models with lower variance, whereas stacking and boosting strive to create robust models with low bias and variation. This research uses the stacking ensemble method to combine multiple models, such as KNN, G.B., and SVM, to produce a single precise model. Trained the proposed ensemble mode to forecast previously trained models’ performance and sent classification information to stacking-based ensemble learners. This section presents an overview of the algorithms employed in the ensemble learning models discussed in this study.

A. Nearest Neighbors (KNN): KNN is a non-parametric, instance-based, supervised machine learning technique for classification and regression applications. KNN generates recommendations based on data points closest to the input data point [35].

B. Gradient Boosting: Machine learning techniques like ensemble learning combine weaker models to create a more potent predictive model. This process involves training base learners, calculating residuals, and combining predictions. A learning rate parameter regulates the contribution of each model to avoid overfitting. Methods such as subsampling and depth control are employed.

C. Support Vector Machine: A binary classification model called the Support Vector Machine (SVM) uses feature sets to predict a hyperplane that categorizes data points. The mathematically defined cost function of the model seeks to locate the plane that best divides the data points of two classes by the most significant margin [36].

The goal of the SVM model, which consists of sigmoid, kernel, Gaussian, and basic linear SVM models, is to locate the plane that divides data points of two classes with the most significant margin in an N-dimensional space. According to the sequential Eqs. (1), (2), and (3) [6].

$J(\theta)=\frac{1}{2} \sum_{k=1}^n \theta_J^2$,      (1)

such that

$\theta^k \mathrm{X}^{(n)} \geq 1, \mathrm{y}^{(n)}=1$,      (2)

$\theta^k X^{(n)} \leq-1, \mathrm{y}^{(n)}=0$     (3)

3.5 Building the model

This study proposes stacking SVM, KNN, and GBoost to train multiple classifiers using extracted features. This study develops an ensemble-based stacking prediction model to categorize text pieces as true or fake news. Researchers developed an ensemble stacking learning algorithm (ESLA) using numerous fundamental categorization models. The design uses k-nearest Neighbors, SVM, and Gradient Boosting. ESLA stacks forecasts for better accuracy, making it unique. A trained meta-model, Ensemble Decision, predicts news article integrity using base model predictions from KNN, SVM, or GBoost for strategic ensemble learning model stacking. Using The strengths of each method improves model performance. When dimensions surpass samples, SVM excels—kernel functions model complex non-linear relationships. Good separation makes SVM robust. SVM ensemble prediction excels in binary categorization. It defines ensemble decision bounds. K-Nearest Neighbors solves basic categorization problems. Data distribution is not assumed, which is important in real-world situations where data may not follow theoretical assumptions. KNN provides more natural, distant ensemble decision-making. It finds local patterns other models overlook. Choose Gradient Boosting for binary, definite, and numerical data. It fits well even when overfitted. Flexibility comes from optimizing any differentiable loss function. GBoost enhances ensembles incrementally. Learn from model errors to improve the ensemble. Every model approaches the issue differently. KNN finds local patterns, SVM defines boundaries, and GBoost enhances predictions. Stacked model predictions produce a final model. Our diverse base models give the final model complete characterizations that capture different data aspects. Combine these models to reduce ensemble bias and variation. Some models’ pros outweigh the cons. Trained the suggested ensemble mode to predict model performance and offered stacking-based ensemble learners classification information.

3.6 Evaluation

Performance evaluation methodologies have long been used in categorization and have become standard performance evaluation measures in related fields. The metrics include accuracy, sensitivity, and specificity, and the equations are as follows [37]:

Precision = TP/TP + FP     (4)

The true positive rate (T.P.) and false positive rate (F.P.) greatly influenced the recall or sensitivity of positive instances.

Recall = TP/TP + FN     (5)

The following equation computes the accuracy, the percentage of accurate predictions, and the false-negative rate (F.N.), which stands for the false-negative rate.

Accuracy = TP + TN/TP + TN + FP + FN     (6)

The term “T.N.” denotes true negative, while “sensitivity” refers to the level of positive records that yield the correct result for every positive record.

Sensitivity = TP/TP + FN     (7)

Particularity refers to the accuracy of accurately arranging positive records out of every single positive record.

Specificity = TN/TN + FP     (8)

The F-measure tool processes a few normal data recovery accuracy and reviews measurements.

F= 2 ∗ precision ∗ recall/precision + recall     (9)

True positive and negative (T.P.) and false positive (F.P.) are used for accurate classification, while false negative and F.N. are used for inaccurate classification. The accuracy of a test depends on its ability to accurately distinguish between fake and real news instances, with sensitivity referring to sensitivity and specificity to specificity. The study calculates prediction accuracy using various classifiers and models, with the suggested Stack model showing high training, testing, sensitivity, and specificity values. Regarding results, the ESLA algorithm performed better than other single classifiers. Table 2 illustrates the performance indicators for the proposed model.

Table 2. The result of the proposed model performance

In-depth subject knowledge and the capacity to identify linguistic discrepancies are prerequisites for manually assessing news. Social media is a source of fake news. The method proposed in this paper is used in real life. These platforms can automatically identify false information by utilizing machine learning algorithms. This aids in quickly identifying and removing misinformation. These algorithms allow internet news aggregators to assess an article’s dependability before posting. Users are guaranteed accurate information as a result. Human fact-checkers in fact-checking organizations benefit from models. Programs can identify fake news using vast amounts of data. For these businesses, this increases scalability and efficiency. These models assist governments in keeping an eye on and reining in bogus news that compromises policy, health, and security. Accurate information is needed in public health emergencies; these algorithms can swiftly spot and remove misinformation. Educators can use these models to teach disinformation and media literacy. They also assist in spotting deceptive patterns and outcomes. More significant advantage: In this study, we investigated using machine learning models and a stacked ensemble to identify fake news items. The data we used in our study came from Twitter tweets and contained news pieces from multiple categories to capture the majority of news instead of neatly classifying political news. This strategy aims to identify linguistic traits that set true news apart from false information. Retrieved various textual features from the articles using the Word2Vec Skip-Gram method, then updated the models with these chosen attributes. They were trained and parameterized to guarantee that the learning models performed at their best. Some models have attained a better degree of accuracy in comparison to others. This study examined various performance measures for each method of machining. Using graph theory and deep learning techniques, the group learners—outperforming the individual learners by 0.963 points on all performance measures—could identify the primary sources disseminating false information. Detecting bogus news in videos in real time is another possible future strategy.