An Empirical Comparison of Ensemble model and Deep Learning Models for Multi-Level Arabic Fake News Classification using JoNewsFake Dataset

An Empirical Comparison of Ensemble model and Deep Learning Models for Multi-Level Arabic Fake News Classification using JoNewsFake Dataset | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article An Empirical Comparison of Ensemble model and Deep Learning Models for Multi-Level Arabic Fake News Classification using JoNewsFake Dataset Noor M. Alkudah, Norisma Idris, Aznul Qalid Md Sabri, Mohammad A. M. Abushariah This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7424709/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 03 Feb, 2026 Read the published version in PeerJ Computer Science → Version 1 posted You are reading this latest preprint version Abstract The public's trust and the integrity of information are at risk by the spread of fake news on Arabic social media platforms. Arabic is still underrepresented in multi-label multi-level fake news detection because of its linguistic complexity and lack of resources, despite significant efforts being made in English. This study provides a full comparison of machine learning and deep learning models for multi-label, multi-level Arabic fake news classification. Using the newly constructed JoNewsFake dataset collected from verified Jordanian news agencies, the models were trained to classify news into main categories, subcategories, and fake/real labels. Experiments included traditional ML classifiers (Random Forest, Extra Trees, and LightGBM) and advanced DL models (CNN + Bi-LSTM, CNN + Bi-GRU, fine-tuned CNN variants, and Transformer-based). Results showed that the Extra Trees classifier outperformed all ML and DL models, achieving F1-scores of 95% (Main Category), 98% (Subcategory), and 95% (Fake/Real). Among the DL models, the transformer-based model yielded the best performance for subcategory classification F1 at 88.2%, while CNN + Bi-GRU achieved 93% accuracy in binary fake/real classification. The study confirms the value of combining syntactic, semantic, and emotional features (Tuned AraBERT) to boost classification performance, particularly for complex subcategory tasks. These findings lay the groundwork for future enhancements in Arabic fake news detection using explainable and scalable models. Physical sciences/Engineering Physical sciences/Mathematics and computing Arabic fake news detection multi-level classification JoNewsFake dataset transformer model deep learning ensemble learning AraBERT embeddings Arabic NLP Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. INTRODUCTION Many people get their news from social media platforms because of their ease of access and instant updates. A lot of people in the Arab world are utilizing social media, for example as their primary source of news instead of newspapers and television. Facebook and Twitter are among the most important sources of instant news in Jordan, especially during political and regional events. According to (Thaher, Saheb, Turabieh, & Chantar, 2021) and (Al-Jalabneh & Safori, 2022), social media is better at getting breaking news out than traditional media. But not all of the information that is spread around is reliable. People often change information on purpose because of economic, political, or societal reasons. This makes content that is not completely true or completely fraudulent (Al-Taie, 2025). This unclear conformity between the true and the false resulted in a lot of fake news on digital platforms. This has caused problems like people not knowing the truth and instability in society. Rumours, hoaxes, satirical pieces, misleading headlines, and political stories are just a few examples of the many kinds or types of information that are part of the larger category of fake news (Rahmanian, 2023). Researchers have found that fake news spreads faster than true news (Bsoul, Qusef, & Abu-Soud, 2022), which has motivated a wave of research into developing automated systems to detect its dissemination (Al-Jalabneh & Safori, 2022). For instance, Google launched the "Google News Initiative" [1] as part of a broader effort to combat misinformation and help users distinguish fake from real content. The task of detecting fake news has high limitations due to the evolving nature of misinformation and the need for good linguistic understanding(Alkhair, Meftouh, Othman, & Smaïli, 2019). A fake news detection model is typically made to classify news articles or posts as either fake or real, often leveraging prior examples of both categories. This classification task may be approached as a binary problem or framed within a more complex multi-class or multi-label context (Khalil, Jarrah, & Aldwairi, 2024). Since 2017, researchers have developed various benchmark datasets such as SANAD, ArSarcasm, and ARACOVID19-MFH, MAWQIF, AraNews, AFND, JoNewsFake FA-KES, and RTAnews fake news datasets to improve model training and evaluation (Salem, Feel, Elbassuoni, Jaber, & Farah, 2019, July), (Al-Salemia, Ayoba, Kendallb, & Noaha, 2019), (Ameur & Aliane, 2021), (Alturayeif, Luqman, & Ahme, 2022)and (Khalil, Jarrah, Aldwairi, & Jaradat, 2022) (Wotaifi & Dhannoon, 2023). While much of the literature has focused on English-language datasets and classification models, Arabic fake news detection remains under-explored (Alturayeif, Luqman, & Ahme, 2022), (Al-Taie, 2025). There aren't any models that can handle Arabic multi-label and multi-level classification, which is a required step to show how news is really organised in the real world(Almuzaini & Azmia, 2022). The linguistic challenges associated with Arabic such as morphological complexity, diglossia, and dialectal variation further complicate the detection process and underscore the need for more tailored modeling approaches (Al-Yahya, Al-Khalifa, Al-Baity, AlSaeed, & Essam, 2021). While Arabic fake news detection has attracted increasing attention, most existing studies focus on binary classification tasks and do not address the hierarchical nature of real-world news. Moreover, few works conduct a comprehensive, unified comparison between ensemble machine learning models and deep learning approaches including transformer-based models under a multi-label, multi-level setup. In addition, there is a lack of studies that fine-tune transformer-based models like AraBERT specifically for Arabic fake news detection, or that integrate semantic, syntactic (POS), and emotional signals into a hybrid representation. This study fills in these gaps by looking at a diverse set of models on a real-world dataset (JoNewsFake), introducing a fine-tuned AraBERT model, and incorporating rich linguistic features for enhanced classification performance across multi-label multilevel classification. This study's main contributions are as follows: We compare several ensemble classifiers (Random Forest, Extra Trees, LightGBM) and deep learning models (CNN+Bi-LSTM, CNN+Bi-GRU, fine-tuned CNN variants, Transformer-based models). We employ a hybrid feature (Tuned AraBERT) representation combining semantic embeddings from AraBERT with syntactic (POS tags) and emotional indicators. We evaluate performance across multi-label, multi-level classification tasks: binary fake/real classification, main category classification, and subcategory classification. We analyze model strengths, limitations, and the role of linguistic features in Arabic fake news classification. This is how this paper has been orgainzed: Section 2 provides an overview of related work in Arabic fake news detection and multi-level multi label text classification. Section 3 outlines the research objectives. Section 4 explains the machine learning and deep learning models used, along with the experimental setup and evaluation metrics. Then Section 5 describes the JoNewsFake dataset, including data collection, filtering, and pre-processing steps followed by experimental setup and evaluation matrices in section 6. Section 7 detailed methodology with data precreation and feature extraction to the preprocessing phase. section 8 presents and analyzes the experimental results and detailed discussion and limitans. Finally, Section 9 concludes the paper and suggests potential directions for future research. 2. RELATED WORKS As social media continues to serve as a dominant source for news dissemination, fake news detection has appeared as a censorious NLP challenge. Early studies primarily tackled fake news as a binary classification task using traditional ML techniques such as SVM, RF, and NB (Khalil, Jarrah, Aldwairi, & Jaradat, 2022). With the rise of deep learning, models like CNN, LSTM and Bi-LSTM were adopted to capture contextual and sequential patterns (Salem, Feel, Elbassuoni, Jaber, & Farah, 2019, July). However, real-world news often carries multiple overlapping themes, motivating researchers to shift towards multi-label and multi-level classification approach. These approaches assign more than one label (e.g., politics, local, economy) and also operate across hierarchical levels (e.g., main category, subcategory, veracity). For example, (SOROUR & ABDELKADER, 2022) proposed a hybrid CNN-LSTM model for multi-class fake news classification. (Liu, Zhao, Liang, Zhang, & Jin, 2022) introduced a recommendation-based architecture that integrates deep features for hierarchical decision-making. In the Arabic context, the majority of fake news detection studies still approach the problem as a task of the binary classification, overlooking its hierarchical and multi-dimensional nature (Hadeel, Mohamed, & Orasan, 2020) (Khalil, Jarrah, & Aldwairi, 2024). Only a few recent works have recognized this limitation. For instance, (Almuzaini & Azmia, 2022) discussed the need for hierarchical models tailored to Arabic’s morphological complexity. (Alturayeif, Luqman, & Ahme, 2022) attempted multi-level stance classification, yet did not fully incorporate fake news detection into this layered approach. While Arabic fake news detection has been increasingly addressed in recently, most existing works rely on binary or single-level classification strategies. (Aljedani, Alotaibi, & Taileb, 2020) and (Alruily, 2021) both talked about how there aren't any strong, multi-label, and hierarchical categorisation models for Arabic text, especially when it comes to real-world news streams. There aren't many datasets that allow multi-label, multi-level classification in Arabic yet, which makes it hard to train and test these kinds of models. This highlights the necessity for resources like JoNewsFake, which address this critical gap. In terms of datasets, only a limited number of Arabic corpora offer partial support for multi-label or multi-level annotation. The AFND dataset (Khalil, Jarrah, Aldwairi, & Jaradat, 2022)provides fake/real labels but is limited to flat classification. FA-KES (Salem, Feel, Elbassuoni, Jaber, & Farah, 2019, July)includes topic-level labels such as politics and religion, while RTAnews (Al-Salemia, Ayoba, Kendallb, & Noaha, 2019) offers multi-topic news articles tagged with several content categories. However, none of these datasets were explicitly designed to support hierarchical or multi-level classification structures. They lack a clear taxonomy of main categories and subcategories, and thus cannot be directly used to train models for multi-level fake news detection without extensive adaptation. This gap significantly limits the development and benchmarking of advanced models in this domain. The research on multi-label and multi-level English-language classification have looked at a bigger range of problem transformation methods, including Label Powerset (LP), and Classifier Chains (CC), as well as models based on ensemble and deep learning. Researchers like (Tarekegn, Ullah, & Cheikh, 2024) and (Yildirim, Bakhshi, & Can, 2024, October) have set out basic ways to turn multi-label issues into multiple binary or multiclass tasks using classifiers like k-NN, Naïve Bayes, and SMO. Other studies, including (Maxwell, et al., 2017)and (Qandos, et al., 2024), used medical or product review datasets to test LSTM and deep neural networks. These studies show that multi-label approach may be used in many different fields. Also, research (Kaur, Kumar, & Kumaraguru, 2020) suggested combining several English datasets (such Reuters and Kaggle) to train multi-level voting classifiers to find fake news. These classifiers worked well in both accuracy and training speed. Therefore, while efforts are emerging, Arabic fake news detection still lacks strong models and resources for multi-label and multi-level classification an area that this study focused to address directly. This study is the first to offer a multi-label, multi-level Arabic fake news classification models, based on JoNewsFake enriched with syntactic and emotional features, enabling deeper experimentation and model evaluation in this under-researched domain. 3. RESEARCH OBJECTIVE Fake news has spread faster as people depend more on social media as their main source of news, especially in Arabic digital media. A lot of early work has been on binary fake news detection, but this doesn't work because real-world news is generally about more than one issue and is organized into hierarchical categories. Moreover, traditional solutions such as blacklisting sources or relying solely on human verification are insufficient, especially when credible sources occasionally disseminate misleading content. Given the linguistic richness and structural complexity of Arabic, a more advanced and context-aware detection approach is needed. Therefore, this research aims to develop a multi-label, multi-level Arabic fake news detection model capable of classifying news articles not only by veracity (fake/real), but also by main and subcategories (e.g., politics → local). The study leverages ML and DL models to learn semantic, syntactic, and emotional patterns that distinguish fake from real news. This work also introduces a novel Arabic dataset specifically designed to support hierarchical classification, addressing an important gap in current fake news detection resources. 4. BACKGROUND This section introduces a concise overview of the ML and DL models employed in this study for multi-label and multi-level Arabic fake news classification across tasks. 4.1 Random Forest (RF) Random Forest is a widely-used ensemble learning model that builds a lot of decision trees while it is training and aggregates their outputs to produce robust predictions. Using a method called bagging, each tree is created on a random group of the training data (bootstrap aggregating), and random subsets of features are selected at each node split to introduce further variance reduction and prevent overfitting. This ensemble strategy enhances both classification accuracy and model generalization, especially in cases involving noisy or complex data (Al-Ahmad, Al-Zoubi, Khurma, & Aljarah, 2021 ) and (Imani, Beikmohammadi, & Arabnia, 2025 ). Several research studies have confirmed the performance of Random Forest in text classification, such as fake news detection. For instance, (ALAZAB, et al., 2022)highlighted its suitability for handling structured and unstructured data with moderate class imbalance. Similarly, (Fouad, Sabbeh, & Medhat, 2022 ) demonstrated its ability to produce stable predictions while maintaining reasonable computational efficiency in natural language processing tasks. In this research, the RF classifier was executed with optimized hyperparameters, including n_estimators, max_depth, and min_samples_split, to suit the characteristics of the Arabic fake news dataset. The model was applied to all three classification levels main category, subcategory, and fake/real using the hybrid feature representation. While it achieved satisfactory accuracy and served as a strong baseline, its capacity to record dependencies in context was limited compared to deep learning-based models. Figure 1 illustrates the Random Forest classifier architecture used in this study. 4.2 Extra Trees Classifier (ETC) The Extra Trees Classifier, is an ensemble learning method that makes a lot of decision trees that haven't been eliminated to enhance prediction accuracy and reduce variance. Extra Trees adds more unpredictability by choosing split points randomly from the range of available feature values. This approach is different from Random Forest, which chooses the optimum and best split based on information gain or the Gini index. It also trains each tree on the whole dataset instead of using bootstrapped samples, which makes training faster and better at generalizing (Patil, 2022 ) (Imran, et al., 2024). Extra Trees is particularly suitable for high-dimensional data, where irrelevant or noisy features may exist. Its robustness lies in its ability to decorrelate individual trees, which helps prevent overfitting. As with other ensemble methods, final decisions are made by aggregating the outputs of all trees in the ensemble. Recent studies have validated the effectiveness of the Extra Trees model in fake news detection. For instance, (Patil, 2022 ) reported a classification accuracy of 94.5%, demonstrating its reliability. Similarly, (Fauzan, Putri, & Laura, 2024 ) found Extra Trees to outperform several models, achieving 96% accuracy when combined with TF-IDF and BoW features alongside PCA and Chi-square selection. These results confirm the model’s strong predictive power when paired with effective feature engineering. In this research, the Extra Trees model was trained using 100 estimators with max_depth = 30 and max_features = 'sqrt'. It was applied to all three classification tasks main category, subcategory, and fake/real using hybrid textual features that included AraBERT embeddings, part-of-speech (POS) tags, and emotional indicators. The model consistently achieved high accuracy, particularly in subcategory classification, and served as a reliable benchmark for comparing traditional ensemble methods with deep learning approaches. Figure 2 illustrates the workflow of the Extra Trees classifier and its architecture that was applied in this research. 4.3 LightGBM (LGBM) LightGBM (Light Gradient Boosting Machine) is a gradient boosting developed by Microsoft that is optimized for capacity, speed, and scalability. It employs histogram-based decision splitting and leaf-wise tree growth, enabling faster convergence and improved accuracy compared to traditional level-wise boosting techniques. LightGBM also supports parallel training and memory-efficient computations, making it particularly effective in handling large-scale, high-dimensional datasets (Belinda, et al., 2024, October). In recent studies, LightGBM has showed better performance in text classification and fake news detection. For example, a hybrid model proposed by (Dhiman, et al., 2024) that combining LightGBM and BERT for fake news detection, achieving new results across multiple datasets. Similarly, (Pillai, 2024 )and (Barua, Rahman, & Joy, 2025 ) highlighted its competitive performance in multi-class and multilingual NLP tasks. In this research, LightGBM was employed to classify Arabic news articles across three tasks: Main Category, Subcategory, and Fake/Real. The model was trained with an 894-dimensional feature representation that included semantic, syntactic, and affective aspects. We set LightGBM up with n_estimators = 100, max_depth = 10, and learning_rate = 0.1. It showed particularly high accuracy in Subcategory classification, thanks to its fast-training speed and generalization capacity across fine-grained labels. However, its binary classification performance (Fake/Real) showed slight sensitivity to class imbalance. Figure 3 illustrates the workflow of LightGBM training in this study, highlighting its role in Arabic news classification across multiple levels. 4.4 CNN + Bi-LSTM The CNN + Bi-LSTM hybrid model is a deep learning architecture that uses both CNN and Bi-LSTM networks to get the best of both of them. CNN layers are good at identifying local features like n-grams and word collocations. Bi-LSTM layers, also find long-range relationships by looking at sequences in both directions. This kind of combination enhances both local and global feature learning, which is essential for understanding Arabic text. There have been a number of studies that show this hybrid design works well for classifying Arabic text (Jain, Gopalani, & Meena, 2025 ). For example, (Fouad, Sabbeh, & Medhat, 2022 ) reported that combining CNN and Bi-LSTM achieved superior accuracy in Arabic fake news detection by taking advantage of both local and contextual representations. Similarly, (Khalil, Jarrah, & Aldwairi, 2024 ) employed a CNN + Bi-LSTM + Attention model for detecting fake content in Arabic news articles and achieved substantial improvements in F1-score compared to standalone models. Additionally, in this research, the hybrid architecture was further enhanced by integrating an Attention mechanism, which lets the model pay attention to the individual elements of the input that are most important. The full model was trained on a unified 894-dimensional input vector that includes semantic, syntactic, and emotional features. It outputs three classification labels: main category, subcategory, and fake/real. This architecture has been widely adopted in Arabic NLP tasks due to its high performance in multi-label classification problems. Its components CNN, Bi-LSTM, and Attention work synergistically to improve both prediction accuracy and interpretability. Figure 4 illustrates the structure of the CNN + Bi-LSTM + Attention model used in this study, highlighting the flow from input features to multi-task outputs. 4.5 CNN + Bi-GRU The CNN + Bi-GRU hybrid model integrates CNN with Bi-GRU, augmented by an Attention mechanism. This architecture uses CNN's capacity to find local n-gram features and stylistic patterns, and Bi-GRU's ability to find contextual reliance in forward and backwards directions together. This gives a more complete picture of Arabic text sequences. Several prior studies have highlighted the effectiveness of this hybrid structure (Aljohani, 2024 ).For instance, (Wotaifi & Dhannoon, 2023 ) showed that combining CNN and Bi-GRU significantly improved performance on Arabic sentiment and fake news detection tasks. (SOROUR & ABDELKADER, 2022 ) also demonstrated that attention-enhanced Bi-GRU models outperformed traditional GRUs, particularly when classifying complex and morphologically rich Arabic text. This model was trained in this study on a unified 894-dimensional feature vector composed of semantic (AraBERT), syntactic (POS tags), and emotional features. The Attention layer dynamically weighed how relevant distinct tokens were, which improved both the accuracy of predictions and the ability to understand the model. To fix the problem of unbalanced categories, the model was improved utilizing standard callbacks (EarlyStopping, ReduceLROnPlateau, ModelCheckpoint) and class weights. Figure 5 illustrates the CNN + Bi-GRU + Attention model employed in this research, showcasing the data flow from multi-dimensional input to the multi-task output layers (main category, subcategory, and fake/real classification). 4.6 Fine-Tuning CNN + Bi-LSTM To improve the performance of the CNN + Bi-LSTM model across all three classification tasks main category, subcategory, and fake/real fine-tuning was employed as a critical optimization phase. At first, the model was trained using frozen base layers. This let the dense layers learn without the problems that could come from changing all the weights at once. When the first convergence was reached, all of the layers were unfrozen, and the model was retrained with a much lower learning rate to improve the generalisation of the whole architecture. The fine-tuning process adopted the Adam optimizer with a learning rate set to 0.00005, along with the sparse categorical cross-entropy loss function, which is appropriate for integer-encoded multi-class targets. The model was trained over 15 epochs using a batch size of 64. To check the training robustness and stop overfitting, several adaptive callbacks were integrated. The ModelCheckpoint callback was used to store the greatest-performing model based on validation accuracy, while ReduceLROnPlateau was employed to dynamically decrease the learning rate if validation performance stopped growing. Additionally, CSVLogger was used to track training history and support resumption in the event of session interruptions a practical requirement when training on platforms like Google Colab. Each version of the model corresponding to main category, subcategory, and fake/real classification was fine-tuned independently using this strategy. The impact of fine-tuning was particularly notable in improving F1-scores and recall metrics, especially for underrepresented classes that previously suffered from overfitting or poor generalization. Figure 6 illustrates the final architecture of the CNN + Bi-LSTM model after fine-tuning. The model takes as input the 894-dimensional vector (AraBERT embeddings + POS + emotion features), which is reshaped and passed through convolutional and recurrent layers. The attention method makes it easier for the model to pay attention to useful and meaning tokens before producing multi-output predictions for the three classification tasks. 4.7 Fine-Tuning CNN + Bi-GRU To further enhance classification performance, the CNN + Bi-GRU hybrid model was fine-tuned independently for each of the three output tasks: main category, subcategory, and fake/real. The fine-tuning phase began by resuming training from the previously saved weights, with all base layers unfrozen and the learning rate carefully reduced to allow more stable gradient updates. Compared to the Bi-LSTM-based fine-tuning strategy, this experiment adopted a slightly modified configuration of callbacks to optimize training stability and reduce the risk of overfitting. Adam optimizer was used with a learning rate of 0.00005 to train the model. The batch size stayed the same at 64 for all runs. Training lasted for up to 15 epochs, although early stopping was used with a patience value of 2. These values meant that the model would cease training as soon as the validation loss stopped getting better. This prevented them from exerting additional effort and reduced the likelihood of the model overfitting. We also used the ReduceLROnPlateau callback to lower the learning rate when performance levels are off, and we turned on ModelCheckpoint to save the weights of the best-performing model based on validation accuracy. Unlike the CNN + Bi-LSTM setup, no CSVLogger was used, and training progress was instead monitored through standard evaluation metrics and visualization of training curves. The CNN + Bi-GRU model demonstrated more stable training behavior, with faster convergence compared to the Bi-LSTM architecture. It maintained strong generalization across the classification tasks and showed particularly efficient learning on the subcategory classification, benefiting from the Bi-GRU's streamlined gating mechanism and contextual sensitivity. Figure 7 illustrates the fine-tuned CNN + Bi-GRU architecture used in this study. Starting from the 894-dimensional input vector, the model extracts local and contextual features using convolutional and bidirectional GRU layers, followed by an attention mechanism. It then produces multi-output predictions for the main category, subcategory, and fake/real tasks. This design effectively balances training efficiency with high classification performance across all outputs. 4.8 Transformer Encoder Block This research created a Transformer Encoder-based architecture to help with the problems of multi-level and multi-label classification in Arabic fake news detection. The self-attention used by transformer Encoder to capture global dependencies over the whole sequence at once, unlike recurrent models like Bi-LSTM and Bi-GRU, which process tokens one at a time. This parallel attention technique is very helpful for Arabic literature since its syntax is complicated, its morphology is rich, and its meaning changes depending on the situation. The model begins by projecting the 894-dimensional input comprising semantic, syntactic, and emotional features into a 128-dimensional space using a dense layer with ReLU activation (Almarashy, Feizi-Derakhshi, & Salehpour, 2023 ) (Alzahrani & Al-yahya, 2024 ). The resulting vector is reshaped to conform to the expected Transformer input format. At the foundation of the model is a transformer Encoder block that includes multi-head self-attention (with two attention heads), a position-wise feedforward layer, residual connections, and both layer normalization and batch normalization to promote training stability and convergence. After encoding, a GlobalAveragePooling1D layer compresses the contextualized sequence representation into a fixed-size vector. After that, a fully connected dense layer with 64 units and ReLU activation, along with a dropout layer (rate = 0.3) to reduce overfitting. The architecture concludes with softmax-activated output layers that provide classification results for each task. Training was conducted using the RMSprop optimizer with a learning rate of 0.0005, and sparse categorical cross-entropy was adopted as the loss function. The training process included several callbacks such as ModelCheckpoint to store the best model, ReduceLROnPlateau to adjust learning rates when validation performance plateaued, and CSVLogger to record training metrics for potential resumption. Class weights were also applied to mitigate the effects of label imbalance, particularly in subcategory classification. The model was trained for up to 15 epochs with a batch size of 256. This Transformer-based design offers a context-aware and scalable way to do multi-task classification, as shown in Fig. 8 . The model can pick up on minor textual signals and tell the difference between complex categories due to the combination of multi-head attention and feedforward sublayers. This makes it especially good at finding fake news in Arabic, which is a language with a lot of structure and variation. 5. DATASET To support the multi-level classification of Arabic fake and real news, this study used a dataset named JoNewsFake, it was collected by researchers themself. The dataset was specifically curated to address the scarcity of large-scale, annotated Arabic corpora for fake news detection particularly those capturing multi-level and multi-label classification structures. JoNewsFake comprises a total of 50,000 Arabic news posts collected from the official Facebook pages of 12 verified Jordanian news agencies. The information includes a lot of different topics, sources, and places where it was published. This shows how hard it may be to understand Arabic digital media in terms of language and ideas. At initially, the process of gathering data found more than 134,000 public posts. These were then filtered by language, length, and relevance. To make sure the dataset was of high quality and consistent, posts with fewer than four terms, posts that weren't in Arabic, duplicates, and posts that weren't about news reporting were all left out. Following this filtering phase, a clean dataset of 50,000 entries was finalized, each representing a unique news post in Modern Standard Arabic or localized Jordanian dialect. Each post in JoNewsFake is annotated with three distinct labels: A Main Category label (22 total classes), such as politics, economy, education, or health. A Subcategory label (75 classes), offering finer-grained classification within each main category (e.g., Education → School). And A Fake/Real label (binary), indicating the factuality of the news post based on expert judgment and cross-validation with official sources. To enhance model learning and feature richness, each instance also includes: A Clean Text column (after advanced Arabic preprocessing). A Part-of-Speech (POS) Tag sequence to support syntactic analysis. And An Emotional Indicator vector, quantifying sentiment and emotional cues within the post. JoNewsFake provides a reliable foundation for evaluating supervised models under multi-task settings. Its design not only supports binary and multi-class classification, but also enables research in multi-label, multi-level classification a structure rarely presents in existing Arabic datasets. Table 1 showcases two representative examples from the dataset. Table 1 examples from the JoNewsFake dataset Text Post Emotin Main Sub R/F مع المقاطعة ربنا ما بينساهم ورزقهم على الله هاد اقل شي ممكن الواحد يقدمه [('', 'مع/NN'), ('', 'المقاطعة/DTNN'), ('', 'ربنا/VBD'), ('', 'ما/WP'), ('', 'بينساهم/VBD'), ('', 'ورزقهم/NNP'), ('', 'على/IN'), ('', 'الله/NNP'), ('', 'هاد/NNP'), ('', 'اقل/JJR'), ('', 'شي/NNP'), ('', 'ممكن/JJ'), ('', 'الواحد/DTNN'), ('', 'يقدمه/VBP')] ['حزن'] Economy boycott Real ابو حماد نعم هناك تصدير للخضار والفواكه لاسرائيل ويوجد تجار معروفين [('', 'ابو/NNP'), ('', 'حماد/NNP'), ('', 'نعم/UH'), ('', 'هناك/RB'), ('', 'تصدير/NN'), ('', 'للخضار/NN'), ('', 'والفواكه/NNP'), ('', 'لاسرائيل/NNP'), ('', 'ويوجد/NNP'), ('', 'تجار/NN'), ('', 'معروفين/JJ')] ['اشمئزاز'] Economy Local Real ارتفاع عدد شهداء القصف الإسرائيلي الذي استهدف سيارة مدنية في رفح جنوبي قطاع غزة إلى الأردن التلفزيون الأردني [('', 'ارتفاع/NN'), ('', 'عدد/NN'), ('', 'شهداء/NN'), ('', 'القصف/DTNN'), ('', 'الإسرائيلي/DTJJ'), ('', 'الذي/WP'), ('', 'استهدف/VBD'), ('', 'سيارة/NN'), ('', 'مدنية/JJ'), ('', 'في/IN'), ('', 'رفح/NNP'), ('', 'جنوبي/NN'), ('', 'قطاع/NN'), ('', 'غزة/NNP'), ('', 'إلى/VBD'), ('', 'الأردن/DTNN'), ('', 'التلفزيون/DTNN'), ('', 'الأردني/DTJJ')] ['حزن'] Politics Israeli-Palestinian Conflict Real 6. EVALUATION 6.1 Experimental Setup We did all of the trials on Google Colab Pro, which gave us access to GPU acceleration and more memory. The models were implemented using Keras with TensorFlow backend, and training was performed on Arabic text inputs represented by 894-dimensional feature vectors. Standard training strategies and callbacks were used to optimize model performance across all classification tasks. 6.2 Evaluation Metrics We use important metrics like average accuracy, precision, average recall, and the F1 score to judge how each classifier works. These are based on equations 1–4. $$\:\mathbf{A}\mathbf{c}\mathbf{c}\mathbf{u}\mathbf{r}\mathbf{a}\mathbf{c}\mathbf{y}=\frac{TP+TN}{TP+FN+TN+FP}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(1\right)$$ $$\:\mathbf{P}\mathbf{r}\mathbf{e}\mathbf{c}\mathbf{i}\mathbf{s}\mathbf{i}\mathbf{o}\mathbf{n}=\frac{TP}{TP+FP}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(2\right)$$ $$\:\mathbf{R}\mathbf{e}\mathbf{c}\mathbf{a}\mathbf{l}\mathbf{l}=\frac{TP}{TP+FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(3\right)$$ $$\:\mathbf{F}1-\mathbf{S}\mathbf{c}\mathbf{o}\mathbf{r}\mathbf{e}=\frac{2.Recall.\:Precision\:}{Recall+Preciosn\:}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(4\right)\:\:\:\:\:\:\:\:$$ 7. METHODOLOGY The study uses a multi-stage methodology to find fake news in Arabic, with an emphasis on a multi-label, multi-level classification approach that is exclusive to the JoNewsFake dataset. The suggested method has four main steps: preparing the data, extracting features, labelling, and modelling. It is meant to deal with both the language problems of Arabic and the difficulties of hierarchical classification. 7.1 Data Preparation JoNewsFake. Is the dataset used in the study, it was collected by researchers themself from 12 verified Jordanian news agencies’ Facebook pages over a 3-month period (December 2023 – February 2024). The collecting method produced more than 134,000 Arabic news posts. These were then filtered based on certain criteria, such as post length, language verification, and metadata availability. This left a clean corpus of 50,000 unique news posts. The filtering process removed irrelevant posts, posts that were too short, duplicate content, or not in Arabic. This procedure made sure that the data set was linguistically sound and semantically rich, making it good for classification tasks later on. Posts were annotated based on three axes: main category, subcategory, and fake/real. A group of four trained annotators put the labels manually, and then two domain experts checked them to ensure they were correct. The team made sure that the annotation process was of excellent quality, with inter-annotator agreement rates exceeding 98%. 7.2 Textual Feature Extraction To improve model performance and support contextual understanding of the Arabic language, we extracted three types of features: AraBERT was used to extract the semantic embeddings and create 768-dimensional contextual embeddings from the [CLS] token in each sentence (Antoun et al., 2020). PoS Tags to Extract the Syntactic Features; where Each word in the text was assigned a Part-of-Speech tag using the Farasa toolkit (Abdelali et al., 2016), which was then converted into a 100-dimensional representation using one-hot or frequency-based encoding. Emotional indicators were extracted, such as the work of Al-Khatib et al. (2020), emotional signals (e.g., joy, anger, sadness) were extracted using an Arabic emotion lexicon and encoded into a 26-dimensional binary vector. All three feature types were concatenated into a unified 894-dimensional input vector per instance, providing a hybrid representation that captures semantic, syntactic, and emotional cues in Arabic news. 7.3 Pre-processing phase To make sure of the quality and consistency of the Arabic text prior to feature extraction, an extensive pre-processing pipeline was applied. The following steps were conducted to clean and standardize the dataset: Non-Arabic characters were removed to ensure linguistic consistency. Diacritic marks, which often vary in informal writing, were stripped to reduce orthographic complexity. Special symbols, punctuation, and URLs were eliminated to minimize noise. Stop words were removed, as they typically carry limited semantic value in classification tasks. Mentions (@user) and other social media artifacts were excluded. We eliminated of duplicate and empty comments, as well as examples that didn't add anything or weren't labelled, to make sure the labels were correct. Stemming also to reduce words to their root, and normalization handled elongated letters and orthographic variations. Prefix and suffix removal was used to minimize morphological redundancy in Arabic text. After applying these cleaning steps, the resulting text was tokenized using the AraBERT pre-trained tokenizer, ensuring compatibility with the downstream semantic embedding process. This comprehensive pre-processing was essential to reduce noise, unify language patterns, and enhance the performance of contextual representation models. 7.4 Handling Class Imbalance Initial data analysis revealed imbalanced class distributions, especially within subcategories and fake/real labels. SMOTE (Synthetic Minority Oversampling Technique) was evaluated but eventually excluded due to poor compatibility with high-dimensional contextual embeddings. Instead, class weighting was applied during model training to counteract imbalance without introducing synthetic noise. 7.5 Overfitting and Cross-Validation Overfitting poses a significant challenge in machine learning, particularly when dealing with deep learning approaches, high-dimensional input characteristics, and imbalanced classes. It means that a model does well on the training data but not on unseen data. To solve this problem, several different solutions were used during the studies. To address this issue, multiple strategies were applied throughout the experiments. To keep overfitting to a minimum in deep learning models, regularisation methods including such as dropout layers, early stopping, and ReduceLROnPlateau were used to minimize overfitting. We also used ModelCheckpoint to select only the best approach performance based on validation accuracy, and CSVLogger to monitor the training progress over time. It is important to highlight that the evaluation of all models was done using a fixed train-test split method (usually 80/20) instead of k-fold cross-validation. This choice was designed to make things easier for computers, especially in deep learning situations where training time and resource needs are considerable. While k-fold cross-validation is often used for more robust performance estimation, the fixed split approach was considered sufficient for fair comparison across all model approaches and classification tasks. 8. EXPERIMENTAL RESULT 8.1 Training and Testing In this study, a fixed 80/20 train-test split strategy was adopted to evaluate the performance of the proposed models. The training set was used to fit each model, while the testing set served for performance evaluation on unseen data. This decision was made to balance computational efficiency with evaluation reliability, especially considering the high training cost of deep learning models. We used Python (v3.8+) and the Keras/TensorFlow approach for deep learning to train all of the models. All of the data were randomly shuffled before being split, using the same seed value for each shuffle. This made sure that the comparisons were fair and could be repeated. We trained each model using the best hyperparameters and regularization methods. The final evaluation looked at accuracy, precision, recall, and F1-score measures. 8.2 Experimental Results To assess the performance of each classifier on Arabic fake news detection, a set of standard evaluation metrics was used, including Accuracy, Precision, Recall, and F1-Score. These metrics were computed across three classification tasks: Main Category (22 classes), Subcategory (75 classes), and Fake/Real (binary). Each model was evaluated using a fixed train-test split (typically 80/20). Table 2 shows a summary of the outcomes of all the models that were compared. Traditional machine learning models like RF and Extra Trees did quite well overall, especially on the Fake/Real classification assignment, where F1-scores were over 94%. Among them, the Extra Trees model achieved the highest consistency across all three classification levels. On the deep learning side, the CNN + Bi-LSTM model initially provided a solid baseline, particularly on Main Category classification. Further enhancement using Bi-GRU and fine-tuned CNN layers yielded improved results. The Transformer Encoder-based model, on the other hand, did better than all the others in Subcategory classification, with an F1-score of about 88.2%. This is because its self-attention mechanism captures long-range relationships, which is especially useful for Arabic, which has a lot of different forms. In general, CNN + Bi-GRU was the best model for Fake/Real classification, with an accuracy rate of over 93%. For Subcategory classification, Transformer Encoder did the best, and for Main Category predictions, Extra Trees stayed strong. These findings indicate that while traditional models remain competitive on simpler tasks, deep learning and transformer-based approach provide superior handling of hierarchical, multi-label text classification problems in Arabic. Figure 3 shows how well all of the models did in the three categorisation tasks: Main Category, Sub Category, and Fake/Real. The Extra Trees classifier was the most accurate of all the models and did well on all tests. Random Forest and LightGBM are two examples of traditional ensemble approaches that also stayed powerful and stable. Traditional ensemble methods such as Random Forest and LightGBM also maintained strong and stable results. Deep learning models like CNN + Bi-LSTM and CNN + Bi-GRU didn't fare very well at first, but after some tweaking, they got a lot better, especially at the Fake/Real classification job. Transformer-based models also fared well since users could see how things were connected in a certain situation. This comparison shows how well ensemble-based models can generalize and how strong they are when working with structured Arabic news data, especially when it comes to hierarchical and multilabel classification problems. Table 2 comparative results of all evaluated models Model Main Category Sub Category Fake/Real Accuracy precision Recall F1 Accuracy precision Recall F1 Accuracy precision Recall F1 RF 0.94 0.94 0.94 0.94 0.98 0.97 0.98 0.97 0.95 0.96 0.94 0.95 Extra Trees 0.95 0.95 0.95 0.95 0.98 0.97 0.98 0.98 0.95 0.96 0.94 0.95 Light GBM 0.94 0.94 0.93 0.94 0.97 0.97 0.97 0.97 0.94 0.94 0.94 0.94 CNN+ Bi-LSTM 0.77 0.76 0.77 0.76 0.89 0.88 0.89 0.88 0.94 0.94 0.94 0.94 CNN + Bi-GRU 0.73 0.73 0.72 0.73 0.89 0.87 0.89 0.88 0.92 0.93 0.92 0.91 Transformer 0.85 0.86 0.85 0.85 0.93 0.93 0.93 0.93 0.94 0.93 0.94 0.92 Tuned CNN + Bi-LSTM 0.79 0.76 0.77 0.79 0.90 0.89 0.88 0.90 0.95 0.94 0.94 0.93 Tuned CNN + Bi-GRU 0.77 0.73 0.73 0.76 0.91 0.90 0.89 0.90 0.94 0.94 0.93 0.94 8.3 Discussion In this study, we performed a comparative analysis of several ML and DL models across three key classification tasks: Main Category, Sub Category, and Fake/Real detection. All models were learned using the same data splits, feature normalization, and balanced class distributions to ensure a fair evaluation. The Extra Trees classifier did the best overall job of all the ML models. It always got high accuracy and F1 ratings on all jobs, but it was especially good at Sub Category categorisation, which had 75 different classes. The model was a good and useful baseline since it could handle biassed information without needing further oversampling; it trained quickly, and it was able to generalise. Similarly, Random Forest and LightGBM also demonstrated strong and stable performance, though they slightly lagged behind Extra Trees in terms of F1-score, especially for the more granular subcategory task. Their ensemble-based nature allowed them to manage complex, high-dimensional feature spaces effectively. The Transformer-based model did better than all the other deep learning models. Because it might understand the context and pay attention to itself, it could capture the hidden dependencies in Arabic language. This is especially useful for hierarchical and multi-label classification tasks. The Transformer model achieved the highest F1-score in the subcategory and fake/real classification, despite requiring significantly more computational resources. Tuned CNN-based models, such as Tuned CNN + Bi-LSTM and Tuned CNN + Bi-GRU, worked better than their untuned variants, especially when it came to detecting fake and real post. They didn't do better than the Transformer model, but they provided a good balance between performance and training speed. In summary, the results confirm that Extra Trees is the most effective traditional ML model for this task, offering both speed and accuracy. Meanwhile, the Transformer-based model proved to be the most powerful DL architecture, particularly suited for the semantic and structural complexity of Arabic fake news classification. 8.4 Limitations and Threats to Validity Although the results of this study appear positive, there are certain problems and potential risks to its validity that should be acknowledged: The dataset comes only from verified Jordanian media sources on Facebook. This restriction makes the results more accurate, but it might be more difficult to use them in other Arabic-speaking areas, on sites like Twitter, or with informal material sources. Ensemble models, such as Extra Trees, and deep models like transformers normally work well but can cause overfitting when used on more variable or noisy datasets. Transformers, in particular, require significant amounts of memory and training time, which may not be feasible in real-time or resource-poor environments. The classification assignments have several degrees of hierarchy (for example, 75 subcategories), which makes learning harder. Some of the wrong classifications may have happened because the meanings of closely related categories were too similar. We only tested the models on data from the same domain. To make sure they can be used in other situations, they should test their performance on outside datasets or in cross-lingual contexts in the future. 8.5 Ethical and Legal Considerations in Arabic Fake News Detection Finding fake news in Arabic media is hard because of technological and language issues, but it also raises big ethical and legal issues. Laws about false information are often unclear in many Arab countries, and they are sometimes used to restrict free speech. When looking for political or religious content, automated detection systems need to be carefully constructed so that they don't make this kind of excessive authority more serious. Also, protecting users' rights and preventing the possible exploitation of detection technology requires making sure that data gathering and model predictions are confidential, truthful, and accessible. 9. CONCLUSION AND FUTURE WORK This study presented a comprehensive evaluation of ML and DL models for Arabic fake news detection using a multi-level, multi-label classification approach. The JoNewsFake dataset, developed from authentic Arabic news sources, served as the foundation for training and testing a range of models across three classification tasks: Main Category, Subcategory, and Fake/Real. Experimental results demonstrated that model performance is significantly influenced by the complexity of the task and the richness of the feature representation. Among ensemble machine learning models, the Extra Trees classifier achieved the highest performance, reaching F1 scores of 95% in the main category, 98% in the subcategory, and 95% in the fake/real classification. On the deep learning side, the Transformer-based model achieved the best overall performance, providing a strong balance across all tasks. It demonstrated superior handling of contextual and hierarchical features, with an F1-score of 88.2% in subcategory classification, approximately 85% in main category, and 94% in fake/real classification. The CNN + Bi-GRU model had the highest accuracy for fake/real classification (93%), but it didn't do as well on the other tasks as the Transformer. These results show how well Transformer models can understand the complicated meanings and structures of Arabic news content. This is the first study to compare ensemble machine learning models, Bi-RNN-based deep learning models, and Transformer-based approache for multi-level, multi-label Arabic fake news detection using real news agency data. In general, the results show how important it is to use scalable and context-aware models to uncover fake news in Arabic. Extra Trees and other traditional models provide good baselines, but transformer-based designs have a lot of promise to improve multi-level Arabic NLP jobs. The findings underscore the effectiveness of ensemble ML models and transformer-based approach in handling complex, imbalanced, and hierarchical Arabic datasets. The integration of semantic (AraBERT), syntactic (PoS), and emotional features also contributed to performance improvements, especially in challenging subcategory classifications. For future research, we could extend this work in several directions. First, adding images, videos, and metadata to multimodal data could make it easier to find false information, especially on social media. Second, using domain-specific data to fine-tune huge pre-trained Arabic transformer models (like MARBERT and QARiB) may help them generalise better and understand context better. Third, using explainable AI (XAI) methods would assist in understanding model predictions, thereby making the systems more transparent and reliable. Also, looking into cross-lingual transfer learning could help Arabic fake news detection models use a lot of resources that are available in English and other languages. Lastly, it would be important for real-world uses to test these models in real-time settings and see how well they hold up with practical scenarios. Declarations Funding Declaration : This research received no external funding. Author Contribution The authors confirms sole responsibility for all aspects of the study, including conceptualization, methodology, data collection, analysis, software implementation, visualization, writing, and final approval of the manuscript. Acknowledgements This dataset and accompanying study are part of the first author's PhD thesis at Universiti Malaya. Data Availability The datasets generated and analysed during the current study are not publicly available in full because they are under review for separate publication. However, a representative sample has been provided as supplementary material, and the complete dataset is available from the corresponding author on reasonable request. References Al-Ahmad, B., Al-Zoubi, A. M., Khurma, R. A., & Aljarah, I. (2021). An Evolutionary Fake News Detection Method for COVID-19 Pandemic Information. Symmetry, 13(6), 1091 . ALAZAB, M., AWAJAN, A., ALAZAB, A., KHREISAT, A., ALHYARI, A., & SAADEH, R. (2022). FAKE-NEWS DETECTION SYSTEM USING MACHINE LEARNING ALGORITHMS FOR ARABIC-LANGUAGE. Journal of Theoretical and Applied Information Technology , 5056-5069. Al-Jalabneh, A. A., & Safori, A. O. (2022). Covid-19 and Misinformation Prevalence: A Content Analysis of Fake News Stories Spread in Jordan. In The Implementation of Smart Technologies for Business Success and Sustainability: During COVID-19 Crises in Developing Countries (pp. 535-545). Cham: Springer International Publishing. Aljedani, N., Alotaibi, R., & Taileb, M. (2020). Multi-Label Arabic Text Classification: An Overview. International Journal of Advanced Computer Science and Applications, 11(10). Aljohani, E. (2024). Enhancing Arabic Fake News Detection: Evaluating Data Balancing Techniques Across Multiple Machine Learning Models. Engineering, Technology & Applied Science Research, , 14(4), 15947-15956. Alkhair, M., Meftouh, K., Othman, N., & Smaïli, K. (2019). An arabic corpus of fake news: Collection, analysis and classification." . Arabic Language Processing: From Theory to Practice: 7th International Conference, ICALP 2019, Nancy, France, October 16–17, 2019, Proceedings 7. Springer International Publishing , pp. 292-302. Almarashy, A. H., Feizi-Derakhshi, M.-R., & Salehpour, P. (2023). Enhancing fake news detection by multi-feature classification. IEEE Access, 11, 139601-139613. Almuzaini, H. A., & Azmia, A. M. (2022). An unsupervised annotation of Arabic texts using multi-label topic modeling and genetic algorithm. Expert Systems with Applications, 203, 117384. Alruily, M. (2021). Classification of Arabic Tweets: A Review. Electronics, 10(10), 1143 . Al-Salemia, B., Ayoba, M., Kendallb, G., & Noaha, S. A. (2019). Multi-label Arabic text categorization: A benchmark and baseline comparison of multi-label learning algorithms. Information Processing & Management, 56(1), 212-227. Al-Taie, M. Z. (2025). Comparative Study of Machine Learning Approaches for Detecting Fake News in Arabic Text. IETI Transactions on Data Analysis and Forecasting, 3(1), , 18-31. Alturayeif, N., Luqman, H., & Ahme, M. (2022). MAWQIF: A Multi-label Arabic Dataset for Target-specific Stance Detection. In Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP) , 174-184. Al-Yahya, M., Al-Khalifa, H., Al-Baity, H., AlSaeed, D., & Essam, A. (2021). Arabic Fake News Detection: Comparative Study of Neural Networks and Transformer-Based Approaches. Complexity, vol. 2021 , Article ID 5516945, 10 pages. Alzahrani, L., & Al-yahya, M. (2024). Pre-Trained Language Model Ensemble for Arabic Fake News Detection. Mathematics, 12(18), , 1-17. Ameur, M. S., & Aliane, H. (2021). ARACOVID19-MFH: ARABIC COVID-19 MULTI-LABEL FAKE NEWS AND HATE SPEECH DETECTION DATASET. Procedia Computer Science, 189, 232-241. Barua, R., Rahman, M. M., & Joy, U. G. (2025). Comparative analysis of Bangla news classification: a study of fake news detection and multiclass classification using BERT and FastText. International Journal of Computers and Applications , 47(5), 475-485. Belinda, M. J., David, S. A., Kannan, E., Naveena, N. R., Rajathi, K., & Mahesh, C. (2024, October). Fake News Detection Using Logistic Regression and LightGBM. In 2024 2nd International Conference on Self Sustainable Artificial Intelligence Systems (ICSSAS) , (pp. (pp. 1-5). IEEE.). Bsoul, M. A., Qusef, A., & Abu-Soud, S. (2022). Building an Optimal Dataset for Arabic Fake News Detection. Procedia Computer Science, 201 , 665-672. Dhiman, P., Kaur, A., Gupta, D., Juneja, S., Nauman, A., & Muhammad, G. (2024). GBERT: A hybrid deep learning model based on GPT-BERT for fake news detection. Heliyon, 10(16). Fauzan, F. J., Putri, C. M., & Laura, P. (2024). Performance Comparison of Classification Algorithms for Chronic Kidney Disease Prediction. IJATIS: Indonesian Journal of Applied Technology and Innovation Science , 1(2), 68-74. Fouad, K. M., Sabbeh, S. F., & Medhat, W. (2022). Arabic Fake News Detection Using Deep Learning. Computers, Materials & Continua, 71(2). Hadeel, S., Mohamed, E., & Orasan, C. (2020). Fake or Real? A Study of Arabic Satirical Fake News. arXiv preprint arXiv:2011.00452 . Imani, M., Beikmohammadi, A., & Arabnia, H. R. (2025). Comprehensive Analysis of Random Forest and XGBoost Performance with SMOTE, ADASYN, and GNUS Under Varying Imbalance Levels. Technologies, 13(3), , 88. Imran, M., Aslam, N., Ahmad, H., Mazhar, F., Bhatti, Y. I., & Abid, U. (2024). Predictive modeling of chronic kidney disease using extra tree classifier: A comparative analysis with traditional methods. Journal of Computing & Biomedical Informatics , 6(02), 261-271. Jain, M. K., Gopalani, D., & Meena, Y. K. (2025). Hybrid CNN-Bi-LSTM model with HHO feature selection for enhanced fake news detection. Social Network Analysis and Mining, 15(1), 43. Kaur, S., Kumar, P., & Kumaraguru, P. (2020). Automating fake news detection system using multi-level voting model. Soft Computing, 24(12), 9049-9069. Khalil, A., Jarrah, M., & Aldwairi, M. (2024). Hybrid Neural Network Models for Detecting Fake News Articles. Human-Centric Intelligent Systems, 4(1) , 136-146. Khalil, A., Jarrah, M., Aldwairi, M., & Jaradat, M. (2022). AFND: Arabic fake news dataset for the detection and classification of articles credibility. Data in Brief . doi:https://doi.org/10.1016/j.dib.2022.108141 Liu, Y. M., Zhao, L., Liang, Y., Zhang, P., & Jin, B. (2022). Hybrid embedding-based text representation for hierarchical multi-label text classification. Expert Systems with Applications, , 187, 115905. Maxwell, A., Li, R., B., Y., Weng, H., Ou, A., Hong, H., & ... & Zhang, C. (2017). Deep learning architectures for multi-label classification of intelligent health risk prediction. BMC bioinformatics, 18 , 121-131. NARRA, M., MUHAMMADUMER, SADIQ, S., ESHMAWI, A., KARAMTI, H., MOHAMED, A., & ASHRAF, I. (2022). Selective Feature Sets Based Fake News Detection for COVID-19 to Manage Infodemic. IEEE Access, 10, , 98724-98736. Nassif, A. B., Elnagar, A., Elgendy, O., & Afadar, Y. (2022). Arabic fake news detection based on deep contextualized embedding models. Neural Computing and Applications, 34(18), , 16019-16032. Patil, D. R. (2022). Fake News Detection Using Majority Voting Technique . arXiv preprint arXiv:2203.09936. Pillai, A. S. (2024). FAKE NEWS MULTI CLASS DETECTION USING TRANSFORMERS AND GRADIENT BOOSTING ENSEMBLE. International Research Journal of Modernization in Engineering Technology and Science , https://www. researchgate. net/publication/378691966_FAKE_NEWS_MULTI_CLAS S_DETECTION_USING_TRANSFORMERS_AND_GRADIENT_BOOSTING_EN SEMBLE. Qandos, N., Hamad, G., Alharb, M., Alturki, S., Alharbi, W., & Albelaihi, A. A. (2024). Multiscale cascaded domain-based approach for Arabic fake reviews detection in e-commerce platforms. Journal of King Saud University-Computer and Information Sciences, 36(2), 101926 . Rahmanian, E. (2023). Fake news: a classification proposal and a future research agenda. Spanish Journal of Marketing-ESIC, 27(1), , 60-78. Salem, F. K., Feel, R. A., Elbassuoni, S., Jaber, M., & Farah, M. (2019, July). Fa-kes: A fake news dataset around the syrian war. In Proceedings of the international AAAI conference on web and social media (Vol. 13, pp. 573-582). SOROUR, S. E., & ABDELKADER, H. E. (2022). AFND: ARABIC FAKE NEWS DETECTION WITH AN ENSEMBLE DEEP CNN-LSTM MODEL. J. Theor. Appl. Inf. Technol, 100(14) , 5072-5086. Tarekegn, A. N., Ullah, M., & Cheikh, F. A. (2024). Deep learning for multi-label learning: A comprehensive survey. arXiv preprint arXiv , 2401.16549. Thaher, T., Saheb, M., Turabieh, H., & Chantar, H. (2021). Intelligent Detection of False Information in Arabic Tweets Utilizing Hybrid Harris Hawks Based Feature Selection and Machine Learning Models. Symmetry, 13(4), 556. Wotaifi, T. A., & Dhannoon, B. N. (2023). An Effective Hybrid Deep Neural Network for Arabic Fake News Detection. Baghdad Science Journal, 20(4) , 1392-1392. Yildirim, O., Bakhshi, S., & Can, F. (2024, October). Prioritized Binary Transformation Method for Efficient Multi-label Classification of Data Streams with Many Labels. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , (pp. (pp. 4218-4222)). Footnotes https://newsinitiative.withgoogle.com/ Additional Declarations No competing interests reported. Supplementary Files JoNewsFakedatasetsample.xlsx Cite Share Download PDF Status: Published Journal Publication published 03 Feb, 2026 Read the published version in PeerJ Computer Science → Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7424709","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":514849721,"identity":"f2b2e728-7ffe-4bea-bfdc-65b493627910","order_by":0,"name":"Noor M. Alkudah","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABFElEQVRIiWNgGAWjYBACfvbG9g8fKmzs+BkYG+CiEvi0SPYcPsY440xasmQDsVoMbqSlMXO2HWbccABJFK8WhgM5Zo8Z2JiZjc8fbv7MU8OQuLaB+eBtHoZtiQ04dDA2nDE3LuBh4zO7kdgmzXOMIXHbAbZkax6G2zi1MDP2GEjPkOBhNrvB2MacwwbSwmMmjU8LGzOPgTSPgQTj5v6DzZ9z/oG08H/Dq4WHjS1NmifBgHEDQ2KDdG4b2BY2vFqATjpsOONAQrIEyC9/+ySMtx1mM7acY3DbGJcW+/sPGx98/Pffjr//+OOPM77ZyG473vzwxpuK27K4tGDYCgwREG3A4EisFiT7SdYxCkbBKBgFwxUAALuIW5B66/PmAAAAAElFTkSuQmCC","orcid":"","institution":"Universiti Malaya","correspondingAuthor":true,"prefix":"","firstName":"Noor","middleName":"M.","lastName":"Alkudah","suffix":""},{"id":514849722,"identity":"6306738b-b7c6-4a6d-9a3d-56a4213cb75f","order_by":1,"name":"Norisma Idris","email":"","orcid":"","institution":"Universiti Malaya","correspondingAuthor":false,"prefix":"","firstName":"Norisma","middleName":"","lastName":"Idris","suffix":""},{"id":514849723,"identity":"fcb145d9-4580-4f55-bfe1-732be93a433c","order_by":2,"name":"Aznul Qalid Md Sabri","email":"","orcid":"","institution":"Universiti Malaya","correspondingAuthor":false,"prefix":"","firstName":"Aznul","middleName":"Qalid Md","lastName":"Sa","suffix":"Md"},{"id":514849724,"identity":"b8dec3b9-81c7-437d-874f-b5f603104c47","order_by":3,"name":"Mohammad A. M. Abushariah","email":"","orcid":"","institution":"The University of Jordan","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"A. M.","lastName":"Abushariah","suffix":""}],"badges":[],"createdAt":"2025-08-21 09:38:33","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7424709/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7424709/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.7717/peerj-cs.3510","type":"published","date":"2026-02-04T00:00:00+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":91459895,"identity":"e270051a-0f5c-4a6f-9eb6-c1f180155dd6","added_by":"auto","created_at":"2025-09-16 16:56:23","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":97296,"visible":true,"origin":"","legend":"\u003cp\u003eRandom Forest structure\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/82a87fd00496a446b1833cb9.png"},{"id":91461354,"identity":"2e38db3d-3655-46e9-9139-84015d0ddb80","added_by":"auto","created_at":"2025-09-16 17:20:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":166914,"visible":true,"origin":"","legend":"\u003cp\u003eExtra Trees Classifier\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/fd0a052872111011811f40ae.png"},{"id":91461201,"identity":"80573d10-66a9-4934-b97c-8932178e4929","added_by":"auto","created_at":"2025-09-16 17:12:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":38113,"visible":true,"origin":"","legend":"\u003cp\u003eThe optimal parameters of the LightGBM model\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/f3fed237b126bd9e22b35320.png"},{"id":91460391,"identity":"bb8329ba-7070-4082-9caa-0c9edb945568","added_by":"auto","created_at":"2025-09-16 17:04:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":227472,"visible":true,"origin":"","legend":"\u003cp\u003eBi-LSTM+CNN (A Hybrid model (\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/efbde467edc479fb7dc8e8aa.png"},{"id":91460390,"identity":"d6c1a263-20e5-4455-ba94-3101603006ab","added_by":"auto","created_at":"2025-09-16 17:04:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":192336,"visible":true,"origin":"","legend":"\u003cp\u003eBi-GRU+CNN (A Hybrid model)\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/0cd674d5a4e16a722353c525.png"},{"id":91459896,"identity":"5a9d3860-7ab7-4a01-a721-b102ccbfd3c8","added_by":"auto","created_at":"2025-09-16 16:56:24","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":240085,"visible":true,"origin":"","legend":"\u003cp\u003eFine-Tuning CNN + Bi-LSTM Architecture\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/856ee4edf31e5f90ebef5a71.png"},{"id":91461355,"identity":"cf944976-9b66-46b6-b977-6c5bc8c2b2f8","added_by":"auto","created_at":"2025-09-16 17:20:24","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":326356,"visible":true,"origin":"","legend":"\u003cp\u003eFine-Tuning CNN + Bi-GRU Architecture\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/a1e916d5736b61f31cb88a07.png"},{"id":91459902,"identity":"86c387d2-a8b2-49a2-8c95-39c2bef50ed2","added_by":"auto","created_at":"2025-09-16 16:56:24","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":124977,"visible":true,"origin":"","legend":"\u003cp\u003eTransformer-based architecture\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/514f5e9b6c290069cea375e5.png"},{"id":91460397,"identity":"665549e6-dfd3-430b-bcd8-491886a41829","added_by":"auto","created_at":"2025-09-16 17:04:24","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":81222,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy Comparison Across Models And Classification Tasks\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/80d09b6edffcb02f7286cb00.png"},{"id":102875272,"identity":"3c970b68-cc04-481d-b805-33bf5b3eef35","added_by":"auto","created_at":"2026-02-17 19:34:15","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2239432,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/60f544ed-1a93-4d3d-aeb5-d1b80832ec8e.pdf"},{"id":91460393,"identity":"de25db2d-495f-4606-bcaf-b92a6843e791","added_by":"auto","created_at":"2025-09-16 17:04:24","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":2328944,"visible":true,"origin":"","legend":"","description":"","filename":"JoNewsFakedatasetsample.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7424709/v1/fa460231a764290d8ebb27b8.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"An Empirical Comparison of Ensemble model and Deep Learning Models for Multi-Level Arabic Fake News Classification using JoNewsFake Dataset","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eMany people get their news from social media platforms because of their ease of access and instant updates. A lot of people in the Arab world are utilizing social media, for example as their primary source of news instead of newspapers and television. Facebook and Twitter are among the most important sources of instant news in Jordan, especially during political and regional events. According to (Thaher, Saheb, Turabieh, \u0026amp; Chantar, 2021) and (Al-Jalabneh \u0026amp; Safori, 2022), social media is better at getting breaking news out than traditional media. But not all of the information that is spread around is reliable. People often change information on purpose because of economic, political, or societal reasons. This makes content that is not completely true or completely fraudulent (Al-Taie, 2025).\u003c/p\u003e\n\u003cp\u003eThis unclear conformity between the true and the false resulted in a lot of fake news on digital platforms. This has caused problems like people not knowing the truth and instability in society. Rumours, hoaxes, satirical pieces, misleading headlines, and political stories are just a few examples of the many kinds or types of information that are part of the larger category of fake news (Rahmanian, 2023). Researchers have found that fake news spreads faster than true news (Bsoul, Qusef, \u0026amp; Abu-Soud, 2022), which has motivated a wave of research into developing automated systems to detect its dissemination (Al-Jalabneh \u0026amp; Safori, 2022). For instance, Google launched the \"Google News Initiative\"\u003csup\u003e[1]\u003c/sup\u003e as part of a broader effort to combat misinformation and help users distinguish fake from real content.\u003c/p\u003e\n\u003cp\u003eThe task of detecting fake news has high limitations due to the evolving nature of misinformation and the need for good linguistic understanding(Alkhair, Meftouh, Othman, \u0026amp; Smaïli, 2019). A fake news detection model is typically made to classify news articles or posts as either fake or real, often leveraging prior examples of both categories. This classification task may be approached as a binary problem or framed within a more complex multi-class or multi-label context (Khalil, Jarrah, \u0026amp; Aldwairi, 2024). Since 2017, researchers have developed various benchmark datasets such as SANAD, ArSarcasm, and ARACOVID19-MFH, MAWQIF, AraNews, AFND, JoNewsFake FA-KES, and RTAnews fake news datasets to improve model training and evaluation (Salem, Feel, Elbassuoni, Jaber, \u0026amp; Farah, 2019, July), (Al-Salemia, Ayoba, Kendallb, \u0026amp; Noaha, 2019), (Ameur \u0026amp; Aliane, 2021), (Alturayeif, Luqman, \u0026amp; Ahme, 2022)and (Khalil, Jarrah, Aldwairi, \u0026amp; Jaradat, 2022) (Wotaifi \u0026amp; Dhannoon, 2023).\u003c/p\u003e\n\u003cp\u003eWhile much of the literature has focused on English-language datasets and classification models, Arabic fake news detection remains under-explored (Alturayeif, Luqman, \u0026amp; Ahme, 2022), (Al-Taie, 2025). There aren't any models that can handle Arabic multi-label and multi-level classification, which is a required step to show how news is really organised in the real world(Almuzaini \u0026amp; Azmia, 2022). The linguistic challenges associated with Arabic such as morphological complexity, diglossia, and dialectal variation further complicate the detection process and underscore the need for more tailored modeling approaches (Al-Yahya, Al-Khalifa, Al-Baity, AlSaeed, \u0026amp; Essam, 2021).\u003c/p\u003e\n\u003cp\u003eWhile Arabic fake news detection has attracted increasing attention, most existing studies focus on binary classification tasks and do not address the hierarchical nature of real-world news. Moreover, few works conduct a comprehensive, unified comparison between ensemble machine learning models and deep learning approaches including transformer-based models under a multi-label, multi-level setup. In addition, there is a lack of studies that fine-tune transformer-based models like AraBERT specifically for Arabic fake news detection, or that integrate semantic, syntactic (POS), and emotional signals into a hybrid representation. This study fills in these gaps by looking at a diverse set of models on a real-world dataset (JoNewsFake), introducing a fine-tuned AraBERT model, and incorporating rich linguistic features for enhanced classification performance across multi-label multilevel classification.\u003c/p\u003e\n\u003cp\u003eThis study's main contributions are as follows:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003eWe compare several ensemble classifiers (Random Forest, Extra Trees, LightGBM) and deep learning models (CNN+Bi-LSTM, CNN+Bi-GRU, fine-tuned CNN variants, Transformer-based models).\u003c/li\u003e\n \u003cli\u003eWe employ a hybrid feature (Tuned AraBERT) representation combining semantic embeddings from AraBERT with syntactic (POS tags) and emotional indicators.\u003c/li\u003e\n \u003cli\u003eWe evaluate performance across multi-label, multi-level classification tasks: binary fake/real classification, main category classification, and subcategory classification.\u003c/li\u003e\n \u003cli\u003eWe analyze model strengths, limitations, and the role of linguistic features in Arabic fake news classification.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThis is how this paper has been orgainzed: Section 2 provides an overview of related work in Arabic fake news detection and multi-level multi label text classification. Section 3 outlines the research objectives. Section 4 explains the machine learning and deep learning models used, along with the experimental setup and evaluation metrics. Then Section 5 describes the JoNewsFake dataset, including data collection, filtering, and pre-processing steps followed by experimental setup and evaluation matrices in section 6. Section 7 detailed methodology with data precreation and feature extraction to the preprocessing phase. section 8 presents and analyzes the experimental results and detailed discussion and limitans. Finally, Section 9 concludes the paper and suggests potential directions for future research.\u003c/p\u003e"},{"header":"2. RELATED WORKS","content":"\u003cp\u003eAs social media continues to serve as a dominant source for news dissemination, fake news detection has appeared as a censorious NLP challenge. Early studies primarily tackled fake news as a binary classification task using traditional ML techniques such as SVM, RF, and NB (Khalil, Jarrah, Aldwairi, \u0026amp; Jaradat, 2022). With the rise of deep learning, models like CNN, LSTM and Bi-LSTM were adopted to capture contextual and sequential patterns (Salem, Feel, Elbassuoni, Jaber, \u0026amp; Farah, 2019, July).\u003c/p\u003e\n\u003cp\u003eHowever, real-world news often carries multiple overlapping themes, motivating researchers to shift towards multi-label and multi-level classification approach. These approaches assign more than one label (e.g., politics, local, economy) and also operate across hierarchical levels (e.g., main category, subcategory, veracity). For example, (SOROUR \u0026amp; ABDELKADER, 2022) proposed a hybrid CNN-LSTM model for multi-class fake news classification. (Liu, Zhao, Liang, Zhang, \u0026amp; Jin, 2022) introduced a recommendation-based architecture that integrates deep features for hierarchical decision-making.\u003c/p\u003e\n\u003cp\u003eIn the Arabic context, the majority of fake news detection studies still approach the problem as a task of the binary classification, overlooking its hierarchical and multi-dimensional nature (Hadeel, Mohamed, \u0026amp; Orasan, 2020)\u0026nbsp;(Khalil, Jarrah, \u0026amp; Aldwairi, 2024). Only a few recent works have recognized this limitation. For instance, (Almuzaini \u0026amp; Azmia, 2022) discussed the need for hierarchical models tailored to Arabic’s morphological complexity. (Alturayeif, Luqman, \u0026amp; Ahme, 2022) attempted multi-level stance classification, yet did not fully incorporate fake news detection into this layered\u0026nbsp;approach.\u003c/p\u003e\n\u003cp\u003eWhile Arabic fake news detection has been increasingly addressed in recently, most existing works rely on binary or single-level classification strategies.\u0026nbsp;(Aljedani, Alotaibi, \u0026amp; Taileb, 2020)\u0026nbsp;and\u0026nbsp;(Alruily, 2021)\u0026nbsp;both talked about how there aren't any strong, multi-label, and hierarchical categorisation models for Arabic text, especially when it comes to real-world news streams. There aren't many datasets that allow multi-label, multi-level classification in Arabic yet, which makes it hard to train and test these kinds of models. This highlights the necessity for resources like JoNewsFake, which address this critical gap.\u003c/p\u003e\n\u003cp\u003eIn terms of datasets, only a limited number of Arabic corpora offer partial support for multi-label or multi-level annotation. The AFND dataset (Khalil, Jarrah, Aldwairi, \u0026amp; Jaradat, 2022)provides fake/real labels but is limited to flat classification. FA-KES (Salem, Feel, Elbassuoni, Jaber, \u0026amp; Farah, 2019, July)includes topic-level labels such as politics and religion, while RTAnews (Al-Salemia, Ayoba, Kendallb, \u0026amp; Noaha, 2019) offers multi-topic news articles tagged with several content categories. However, none of these datasets were explicitly designed to support hierarchical or multi-level classification structures. They lack a clear taxonomy of main categories and subcategories, and thus cannot be directly used to train models for multi-level fake news detection without extensive adaptation. This gap significantly limits the development and benchmarking of advanced models in this domain.\u003c/p\u003e\n\u003cp\u003eThe research on multi-label and multi-level English-language classification have looked at a bigger range of problem transformation methods, including Label Powerset (LP), and Classifier Chains (CC), as well as models based on ensemble and deep learning. Researchers like (Tarekegn, Ullah, \u0026amp; Cheikh, 2024) and (Yildirim, Bakhshi, \u0026amp; Can, 2024, October) have set out basic ways to turn multi-label issues into multiple binary or multiclass tasks using classifiers like k-NN, Naïve Bayes, and SMO. Other studies, including (Maxwell, et al., 2017)and (Qandos, et al., 2024), used medical or product review datasets to test LSTM and deep neural networks. These studies show that multi-label\u0026nbsp;approach\u0026nbsp;may be used in many different fields. Also, research (Kaur, Kumar, \u0026amp; Kumaraguru, 2020) suggested combining several English datasets (such Reuters and Kaggle) to train multi-level voting classifiers to find fake news. These classifiers worked well in both accuracy and training speed.\u003c/p\u003e\n\u003cp\u003eTherefore, while efforts are emerging, Arabic fake news detection still lacks strong models and resources for multi-label and multi-level classification an area that this study focused to address directly. This study is the first to offer a multi-label, multi-level Arabic fake news classification models, based on JoNewsFake enriched with syntactic and emotional features, enabling deeper experimentation and model evaluation in this under-researched domain.\u003c/p\u003e"},{"header":"3. RESEARCH OBJECTIVE","content":"\u003cp\u003eFake news has spread faster as people depend more on social media as their main source of news, especially in Arabic digital media. A lot of early work has been on binary fake news detection, but this doesn't work because real-world news is generally about more than one issue and is organized into hierarchical categories. Moreover, traditional solutions such as blacklisting sources or relying solely on human verification are insufficient, especially when credible sources occasionally disseminate misleading content. Given the linguistic richness and structural complexity of Arabic, a more advanced and context-aware detection approach is needed. Therefore, this research aims to develop a multi-label, multi-level Arabic fake news detection model capable of classifying news articles not only by veracity (fake/real), but also by main and subcategories (e.g., politics → local). The study leverages ML and DL models to learn semantic, syntactic, and emotional patterns that distinguish fake from real news. This work also introduces a novel Arabic dataset specifically designed to support hierarchical classification, addressing an important gap in current fake news detection resources.\u003c/p\u003e"},{"header":"4. BACKGROUND","content":"\u003cp\u003eThis section introduces a concise overview of the ML and DL models employed in this study for multi-label and multi-level Arabic fake news classification across tasks.\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Random Forest (RF)\u003c/h2\u003e\u003cp\u003eRandom Forest is a widely-used ensemble learning model that builds a lot of decision trees while it is training and aggregates their outputs to produce robust predictions. Using a method called bagging, each tree is created on a random group of the training data (bootstrap aggregating), and random subsets of features are selected at each node split to introduce further variance reduction and prevent overfitting. This ensemble strategy enhances both classification accuracy and model generalization, especially in cases involving noisy or complex data (Al-Ahmad, Al-Zoubi, Khurma, \u0026amp; Aljarah, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) and (Imani, Beikmohammadi, \u0026amp; Arabnia, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eSeveral research studies have confirmed the performance of Random Forest in text classification, such as fake news detection. For instance, (ALAZAB, et al., 2022)highlighted its suitability for handling structured and unstructured data with moderate class imbalance. Similarly, (Fouad, Sabbeh, \u0026amp; Medhat, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) demonstrated its ability to produce stable predictions while maintaining reasonable computational efficiency in natural language processing tasks. In this research, the RF classifier was executed with optimized hyperparameters, including n_estimators, max_depth, and min_samples_split, to suit the characteristics of the Arabic fake news dataset. The model was applied to all three classification levels main category, subcategory, and fake/real using the hybrid feature representation. While it achieved satisfactory accuracy and served as a strong baseline, its capacity to record dependencies in context was limited compared to deep learning-based models. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the Random Forest classifier architecture used in this study.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Extra Trees Classifier (ETC)\u003c/h2\u003e\u003cp\u003eThe Extra Trees Classifier, is an ensemble learning method that makes a lot of decision trees that haven't been eliminated to enhance prediction accuracy and reduce variance. Extra Trees adds more unpredictability by choosing split points randomly from the range of available feature values. This approach is different from Random Forest, which chooses the optimum and best split based on information gain or the Gini index. It also trains each tree on the whole dataset instead of using bootstrapped samples, which makes training faster and better at generalizing (Patil, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) (Imran, et al., 2024). Extra Trees is particularly suitable for high-dimensional data, where irrelevant or noisy features may exist. Its robustness lies in its ability to decorrelate individual trees, which helps prevent overfitting. As with other ensemble methods, final decisions are made by aggregating the outputs of all trees in the ensemble.\u003c/p\u003e\u003cp\u003eRecent studies have validated the effectiveness of the Extra Trees model in fake news detection. For instance, (Patil, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) reported a classification accuracy of 94.5%, demonstrating its reliability. Similarly, (Fauzan, Putri, \u0026amp; Laura, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) found Extra Trees to outperform several models, achieving 96% accuracy when combined with TF-IDF and BoW features alongside PCA and Chi-square selection. These results confirm the model\u0026rsquo;s strong predictive power when paired with effective feature engineering. In this research, the Extra Trees model was trained using 100 estimators with max_depth\u0026thinsp;=\u0026thinsp;30 and max_features = 'sqrt'. It was applied to all three classification tasks main category, subcategory, and fake/real using hybrid textual features that included AraBERT embeddings, part-of-speech (POS) tags, and emotional indicators. The model consistently achieved high accuracy, particularly in subcategory classification, and served as a reliable benchmark for comparing traditional ensemble methods with deep learning approaches. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the workflow of the Extra Trees classifier and its architecture that was applied in this research.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e4.3 LightGBM (LGBM)\u003c/h2\u003e\u003cp\u003eLightGBM (Light Gradient Boosting Machine) is a gradient boosting developed by Microsoft that is optimized for capacity, speed, and scalability. It employs histogram-based decision splitting and leaf-wise tree growth, enabling faster convergence and improved accuracy compared to traditional level-wise boosting techniques. LightGBM also supports parallel training and memory-efficient computations, making it particularly effective in handling large-scale, high-dimensional datasets (Belinda, et al., 2024, October). In recent studies, LightGBM has showed better performance in text classification and fake news detection. For example, a hybrid model proposed by (Dhiman, et al., 2024) that combining LightGBM and BERT for fake news detection, achieving new results across multiple datasets. Similarly, (Pillai, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)and (Barua, Rahman, \u0026amp; Joy, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) highlighted its competitive performance in multi-class and multilingual NLP tasks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn this research, LightGBM was employed to classify Arabic news articles across three tasks: Main Category, Subcategory, and Fake/Real. The model was trained with an 894-dimensional feature representation that included semantic, syntactic, and affective aspects. We set LightGBM up with n_estimators\u0026thinsp;=\u0026thinsp;100, max_depth\u0026thinsp;=\u0026thinsp;10, and learning_rate\u0026thinsp;=\u0026thinsp;0.1. It showed particularly high accuracy in Subcategory classification, thanks to its fast-training speed and generalization capacity across fine-grained labels. However, its binary classification performance (Fake/Real) showed slight sensitivity to class imbalance. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates the workflow of LightGBM training in this study, highlighting its role in Arabic news classification across multiple levels.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e4.4 CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM\u003c/h2\u003e\u003cp\u003eThe CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM hybrid model is a deep learning architecture that uses both CNN and Bi-LSTM networks to get the best of both of them. CNN layers are good at identifying local features like n-grams and word collocations. Bi-LSTM layers, also find long-range relationships by looking at sequences in both directions. This kind of combination enhances both local and global feature learning, which is essential for understanding Arabic text. There have been a number of studies that show this hybrid design works well for classifying Arabic text (Jain, Gopalani, \u0026amp; Meena, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). For example, (Fouad, Sabbeh, \u0026amp; Medhat, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) reported that combining CNN and Bi-LSTM achieved superior accuracy in Arabic fake news detection by taking advantage of both local and contextual representations. Similarly, (Khalil, Jarrah, \u0026amp; Aldwairi, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) employed a CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM\u0026thinsp;+\u0026thinsp;Attention model for detecting fake content in Arabic news articles and achieved substantial improvements in F1-score compared to standalone models. Additionally, in this research, the hybrid architecture was further enhanced by integrating an Attention mechanism, which lets the model pay attention to the individual elements of the input that are most important. The full model was trained on a unified 894-dimensional input vector that includes semantic, syntactic, and emotional features. It outputs three classification labels: main category, subcategory, and fake/real. This architecture has been widely adopted in Arabic NLP tasks due to its high performance in multi-label classification problems. Its components CNN, Bi-LSTM, and Attention work synergistically to improve both prediction accuracy and interpretability. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the structure of the CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM\u0026thinsp;+\u0026thinsp;Attention model used in this study, highlighting the flow from input features to multi-task outputs.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e4.5 CNN\u0026thinsp;+\u0026thinsp;Bi-GRU\u003c/h2\u003e\u003cp\u003eThe CNN\u0026thinsp;+\u0026thinsp;Bi-GRU hybrid model integrates CNN with Bi-GRU, augmented by an Attention mechanism. This architecture uses CNN's capacity to find local n-gram features and stylistic patterns, and Bi-GRU's ability to find contextual reliance in forward and backwards directions together. This gives a more complete picture of Arabic text sequences. Several prior studies have highlighted the effectiveness of this hybrid structure (Aljohani, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).For instance, (Wotaifi \u0026amp; Dhannoon, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) showed that combining CNN and Bi-GRU significantly improved performance on Arabic sentiment and fake news detection tasks. (SOROUR \u0026amp; ABDELKADER, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) also demonstrated that attention-enhanced Bi-GRU models outperformed traditional GRUs, particularly when classifying complex and morphologically rich Arabic text.\u003c/p\u003e\u003cp\u003eThis model was trained in this study on a unified 894-dimensional feature vector composed of semantic (AraBERT), syntactic (POS tags), and emotional features. The Attention layer dynamically weighed how relevant distinct tokens were, which improved both the accuracy of predictions and the ability to understand the model. To fix the problem of unbalanced categories, the model was improved utilizing standard callbacks (EarlyStopping, ReduceLROnPlateau, ModelCheckpoint) and class weights. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e illustrates the CNN\u0026thinsp;+\u0026thinsp;Bi-GRU\u0026thinsp;+\u0026thinsp;Attention model employed in this research, showcasing the data flow from multi-dimensional input to the multi-task output layers (main category, subcategory, and fake/real classification).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e4.6 Fine-Tuning CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM\u003c/h2\u003e\u003cp\u003eTo improve the performance of the CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM model across all three classification tasks main category, subcategory, and fake/real fine-tuning was employed as a critical optimization phase. At first, the model was trained using frozen base layers. This let the dense layers learn without the problems that could come from changing all the weights at once. When the first convergence was reached, all of the layers were unfrozen, and the model was retrained with a much lower learning rate to improve the generalisation of the whole architecture. The fine-tuning process adopted the Adam optimizer with a learning rate set to 0.00005, along with the sparse categorical cross-entropy loss function, which is appropriate for integer-encoded multi-class targets. The model was trained over 15 epochs using a batch size of 64. To check the training robustness and stop overfitting, several adaptive callbacks were integrated. The ModelCheckpoint callback was used to store the greatest-performing model based on validation accuracy, while ReduceLROnPlateau was employed to dynamically decrease the learning rate if validation performance stopped growing. Additionally, CSVLogger was used to track training history and support resumption in the event of session interruptions a practical requirement when training on platforms like Google Colab. Each version of the model corresponding to main category, subcategory, and fake/real classification was fine-tuned independently using this strategy. The impact of fine-tuning was particularly notable in improving F1-scores and recall metrics, especially for underrepresented classes that previously suffered from overfitting or poor generalization. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates the final architecture of the CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM model after fine-tuning. The model takes as input the 894-dimensional vector (AraBERT embeddings\u0026thinsp;+\u0026thinsp;POS\u0026thinsp;+\u0026thinsp;emotion features), which is reshaped and passed through convolutional and recurrent layers. The attention method makes it easier for the model to pay attention to useful and meaning tokens before producing multi-output predictions for the three classification tasks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e4.7 Fine-Tuning CNN\u0026thinsp;+\u0026thinsp;Bi-GRU\u003c/h2\u003e\u003cp\u003eTo further enhance classification performance, the CNN\u0026thinsp;+\u0026thinsp;Bi-GRU hybrid model was fine-tuned independently for each of the three output tasks: main category, subcategory, and fake/real. The fine-tuning phase began by resuming training from the previously saved weights, with all base layers unfrozen and the learning rate carefully reduced to allow more stable gradient updates. Compared to the Bi-LSTM-based fine-tuning strategy, this experiment adopted a slightly modified configuration of callbacks to optimize training stability and reduce the risk of overfitting.\u003c/p\u003e\u003cp\u003eAdam optimizer was used with a learning rate of 0.00005 to train the model. The batch size stayed the same at 64 for all runs. Training lasted for up to 15 epochs, although early stopping was used with a patience value of 2. These values meant that the model would cease training as soon as the validation loss stopped getting better. This prevented them from exerting additional effort and reduced the likelihood of the model overfitting. We also used the ReduceLROnPlateau callback to lower the learning rate when performance levels are off, and we turned on ModelCheckpoint to save the weights of the best-performing model based on validation accuracy. Unlike the CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM setup, no CSVLogger was used, and training progress was instead monitored through standard evaluation metrics and visualization of training curves. The CNN\u0026thinsp;+\u0026thinsp;Bi-GRU model demonstrated more stable training behavior, with faster convergence compared to the Bi-LSTM architecture. It maintained strong generalization across the classification tasks and showed particularly efficient learning on the subcategory classification, benefiting from the Bi-GRU's streamlined gating mechanism and contextual sensitivity.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e illustrates the fine-tuned CNN\u0026thinsp;+\u0026thinsp;Bi-GRU architecture used in this study. Starting from the 894-dimensional input vector, the model extracts local and contextual features using convolutional and bidirectional GRU layers, followed by an attention mechanism. It then produces multi-output predictions for the main category, subcategory, and fake/real tasks. This design effectively balances training efficiency with high classification performance across all outputs.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e4.8 Transformer Encoder Block\u003c/h2\u003e\u003cp\u003eThis research created a Transformer Encoder-based architecture to help with the problems of multi-level and multi-label classification in Arabic fake news detection. The self-attention used by transformer Encoder to capture global dependencies over the whole sequence at once, unlike recurrent models like Bi-LSTM and Bi-GRU, which process tokens one at a time. This parallel attention technique is very helpful for Arabic literature since its syntax is complicated, its morphology is rich, and its meaning changes depending on the situation. The model begins by projecting the 894-dimensional input comprising semantic, syntactic, and emotional features into a 128-dimensional space using a dense layer with ReLU activation (Almarashy, Feizi-Derakhshi, \u0026amp; Salehpour, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) (Alzahrani \u0026amp; Al-yahya, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The resulting vector is reshaped to conform to the expected Transformer input format. At the foundation of the model is a transformer Encoder block that includes multi-head self-attention (with two attention heads), a position-wise feedforward layer, residual connections, and both layer normalization and batch normalization to promote training stability and convergence.\u003c/p\u003e\u003cp\u003eAfter encoding, a GlobalAveragePooling1D layer compresses the contextualized sequence representation into a fixed-size vector. After that, a fully connected dense layer with 64 units and ReLU activation, along with a dropout layer (rate\u0026thinsp;=\u0026thinsp;0.3) to reduce overfitting. The architecture concludes with softmax-activated output layers that provide classification results for each task. Training was conducted using the RMSprop optimizer with a learning rate of 0.0005, and sparse categorical cross-entropy was adopted as the loss function. The training process included several callbacks such as ModelCheckpoint to store the best model, ReduceLROnPlateau to adjust learning rates when validation performance plateaued, and CSVLogger to record training metrics for potential resumption. Class weights were also applied to mitigate the effects of label imbalance, particularly in subcategory classification. The model was trained for up to 15 epochs with a batch size of 256. This Transformer-based design offers a context-aware and scalable way to do multi-task classification, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e. The model can pick up on minor textual signals and tell the difference between complex categories due to the combination of multi-head attention and feedforward sublayers. This makes it especially good at finding fake news in Arabic, which is a language with a lot of structure and variation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"5. DATASET","content":"\u003cp\u003eTo support the multi-level classification of Arabic fake and real news, this study used a dataset named JoNewsFake, it was collected by researchers themself. The dataset was specifically curated to address the scarcity of large-scale, annotated Arabic corpora for fake news detection particularly those capturing multi-level and multi-label classification structures. JoNewsFake comprises a total of 50,000 Arabic news posts collected from the official Facebook pages of 12 verified Jordanian news agencies. The information includes a lot of different topics, sources, and places where it was published. This shows how hard it may be to understand Arabic digital media in terms of language and ideas. At initially, the process of gathering data found more than 134,000 public posts. These were then filtered by language, length, and relevance. To make sure the dataset was of high quality and consistent, posts with fewer than four terms, posts that weren't in Arabic, duplicates, and posts that weren't about news reporting were all left out. Following this filtering phase, a clean dataset of 50,000 entries was finalized, each representing a unique news post in Modern Standard Arabic or localized Jordanian dialect.\u003c/p\u003e\u003cp\u003eEach post in JoNewsFake is annotated with three distinct labels: A Main Category label (22 total classes), such as politics, economy, education, or health. A Subcategory label (75 classes), offering finer-grained classification within each main category (e.g., Education \u0026rarr; School). And A Fake/Real label (binary), indicating the factuality of the news post based on expert judgment and cross-validation with official sources.\u003c/p\u003e\u003cp\u003eTo enhance model learning and feature richness, each instance also includes: A Clean Text column (after advanced Arabic preprocessing). A Part-of-Speech (POS) Tag sequence to support syntactic analysis. And An Emotional Indicator vector, quantifying sentiment and emotional cues within the post. JoNewsFake provides a reliable foundation for evaluating supervised models under multi-task settings. Its design not only supports binary and multi-class classification, but also enables research in multi-label, multi-level classification a structure rarely presents in existing Arabic datasets. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e showcases two representative examples from the dataset.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eexamples from the JoNewsFake dataset\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eText\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePost\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eEmotin\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMain\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSub\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eR/F\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eمع المقاطعة ربنا ما بينساهم ورزقهم على الله هاد اقل شي ممكن الواحد يقدمه\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e[('', 'مع/NN'), ('', 'المقاطعة/DTNN'), ('', 'ربنا/VBD'), ('', 'ما/WP'), ('', 'بينساهم/VBD'), ('', 'ورزقهم/NNP'), ('', 'على/IN'), ('', 'الله/NNP'), ('', 'هاد/NNP'), ('', 'اقل/JJR'), ('', 'شي/NNP'), ('', 'ممكن/JJ'), ('', 'الواحد/DTNN'), ('', 'يقدمه/VBP')]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e['حزن']\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eEconomy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eboycott\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eReal\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eابو حماد نعم هناك تصدير للخضار والفواكه لاسرائيل ويوجد تجار معروفين\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e[('', 'ابو/NNP'), ('', 'حماد/NNP'), ('', 'نعم/UH'), ('', 'هناك/RB'), ('', 'تصدير/NN'), ('', 'للخضار/NN'), ('', 'والفواكه/NNP'), ('', 'لاسرائيل/NNP'), ('', 'ويوجد/NNP'), ('', 'تجار/NN'), ('', 'معروفين/JJ')]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e['اشمئزاز']\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eEconomy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eLocal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eReal\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eارتفاع عدد شهداء القصف الإسرائيلي الذي استهدف سيارة مدنية في رفح جنوبي قطاع غزة إلى الأردن التلفزيون الأردني\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e[('', 'ارتفاع/NN'), ('', 'عدد/NN'), ('', 'شهداء/NN'), ('', 'القصف/DTNN'), ('', 'الإسرائيلي/DTJJ'), ('', 'الذي/WP'), ('', 'استهدف/VBD'), ('', 'سيارة/NN'), ('', 'مدنية/JJ'), ('', 'في/IN'), ('', 'رفح/NNP'), ('', 'جنوبي/NN'), ('', 'قطاع/NN'), ('', 'غزة/NNP'), ('', 'إلى/VBD'), ('', 'الأردن/DTNN'), ('', 'التلفزيون/DTNN'), ('', 'الأردني/DTJJ')]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e['حزن']\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePolitics\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eIsraeli-Palestinian Conflict\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eReal\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"6. EVALUATION","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e6.1 Experimental Setup\u003c/h2\u003e\u003cp\u003eWe did all of the trials on Google Colab Pro, which gave us access to GPU acceleration and more memory. The models were implemented using Keras with TensorFlow backend, and training was performed on Arabic text inputs represented by 894-dimensional feature vectors. Standard training strategies and callbacks were used to optimize model performance across all classification tasks.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e6.2 Evaluation Metrics\u003c/h2\u003e\u003cp\u003eWe use important metrics like average accuracy, precision, average recall, and the F1 score to judge how each classifier works. These are based on equations 1\u0026ndash;4.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\mathbf{A}\\mathbf{c}\\mathbf{c}\\mathbf{u}\\mathbf{r}\\mathbf{a}\\mathbf{c}\\mathbf{y}=\\frac{TP+TN}{TP+FN+TN+FP}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\mathbf{P}\\mathbf{r}\\mathbf{e}\\mathbf{c}\\mathbf{i}\\mathbf{s}\\mathbf{i}\\mathbf{o}\\mathbf{n}=\\frac{TP}{TP+FP}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\mathbf{R}\\mathbf{e}\\mathbf{c}\\mathbf{a}\\mathbf{l}\\mathbf{l}=\\frac{TP}{TP+FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\mathbf{F}1-\\mathbf{S}\\mathbf{c}\\mathbf{o}\\mathbf{r}\\mathbf{e}=\\frac{2.Recall.\\:Precision\\:}{Recall+Preciosn\\:}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(4\\right)\\:\\:\\:\\:\\:\\:\\:\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"7. METHODOLOGY","content":"\u003cp\u003eThe study uses a multi-stage methodology to find fake news in Arabic, with an emphasis on a multi-label, multi-level classification approach that is exclusive to the JoNewsFake dataset. The suggested method has four main steps: preparing the data, extracting features, labelling, and modelling. It is meant to deal with both the language problems of Arabic and the difficulties of hierarchical classification.\u003c/p\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e7.1 Data Preparation\u003c/h2\u003e\u003cp\u003eJoNewsFake. Is the dataset used in the study, it was collected by researchers themself from 12 verified Jordanian news agencies\u0026rsquo; Facebook pages over a 3-month period (December 2023 \u0026ndash; February 2024). The collecting method produced more than 134,000 Arabic news posts. These were then filtered based on certain criteria, such as post length, language verification, and metadata availability. This left a clean corpus of 50,000 unique news posts.\u003c/p\u003e\u003cp\u003eThe filtering process removed irrelevant posts, posts that were too short, duplicate content, or not in Arabic. This procedure made sure that the data set was linguistically sound and semantically rich, making it good for classification tasks later on. Posts were annotated based on three axes: main category, subcategory, and fake/real. A group of four trained annotators put the labels manually, and then two domain experts checked them to ensure they were correct. The team made sure that the annotation process was of excellent quality, with inter-annotator agreement rates exceeding 98%.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e7.2 Textual Feature Extraction\u003c/h2\u003e\u003cp\u003eTo improve model performance and support contextual understanding of the Arabic language, we extracted three types of features:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eAraBERT was used to extract the semantic embeddings and create 768-dimensional contextual embeddings from the [CLS] token in each sentence (Antoun et al., 2020).\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003ePoS Tags to Extract the Syntactic Features; where Each word in the text was assigned a Part-of-Speech tag using the Farasa toolkit (Abdelali et al., 2016), which was then converted into a 100-dimensional representation using one-hot or frequency-based encoding.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eEmotional indicators were extracted, such as the work of Al-Khatib et al. (2020), emotional signals (e.g., joy, anger, sadness) were extracted using an Arabic emotion lexicon and encoded into a 26-dimensional binary vector.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eAll three feature types were concatenated into a unified 894-dimensional input vector per instance, providing a hybrid representation that captures semantic, syntactic, and emotional cues in Arabic news.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e7.3 Pre-processing phase\u003c/h2\u003e\u003cp\u003eTo make sure of the quality and consistency of the Arabic text prior to feature extraction, an extensive pre-processing pipeline was applied. The following steps were conducted to clean and standardize the dataset:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eNon-Arabic characters were removed to ensure linguistic consistency.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDiacritic marks, which often vary in informal writing, were stripped to reduce orthographic complexity.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSpecial symbols, punctuation, and URLs were eliminated to minimize noise.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eStop words were removed, as they typically carry limited semantic value in classification tasks.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eMentions (@user) and other social media artifacts were excluded.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eWe eliminated of duplicate and empty comments, as well as examples that didn't add anything or weren't labelled, to make sure the labels were correct.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eStemming also to reduce words to their root, and normalization handled elongated letters and orthographic variations.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePrefix and suffix removal was used to minimize morphological redundancy in Arabic text.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eAfter applying these cleaning steps, the resulting text was tokenized using the AraBERT pre-trained tokenizer, ensuring compatibility with the downstream semantic embedding process. This comprehensive pre-processing was essential to reduce noise, unify language patterns, and enhance the performance of contextual representation models.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003e7.4 Handling Class Imbalance\u003c/h2\u003e\u003cp\u003eInitial data analysis revealed imbalanced class distributions, especially within subcategories and fake/real labels. SMOTE (Synthetic Minority Oversampling Technique) was evaluated but eventually excluded due to poor compatibility with high-dimensional contextual embeddings. Instead, class weighting was applied during model training to counteract imbalance without introducing synthetic noise.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003e7.5 Overfitting and Cross-Validation\u003c/h2\u003e\u003cp\u003eOverfitting poses a significant challenge in machine learning, particularly when dealing with deep learning approaches, high-dimensional input characteristics, and imbalanced classes. It means that a model does well on the training data but not on unseen data. To solve this problem, several different solutions were used during the studies. To address this issue, multiple strategies were applied throughout the experiments. To keep overfitting to a minimum in deep learning models, regularisation methods including such as dropout layers, early stopping, and ReduceLROnPlateau were used to minimize overfitting. We also used ModelCheckpoint to select only the best approach performance based on validation accuracy, and CSVLogger to monitor the training progress over time. It is important to highlight that the evaluation of all models was done using a fixed train-test split method (usually 80/20) instead of k-fold cross-validation. This choice was designed to make things easier for computers, especially in deep learning situations where training time and resource needs are considerable. While k-fold cross-validation is often used for more robust performance estimation, the fixed split approach was considered sufficient for fair comparison across all model approaches and classification tasks.\u003c/p\u003e\u003c/div\u003e"},{"header":"8. EXPERIMENTAL RESULT","content":"\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003e8.1 Training and Testing\u003c/h2\u003e\u003cp\u003eIn this study, a fixed 80/20 train-test split strategy was adopted to evaluate the performance of the proposed models. The training set was used to fit each model, while the testing set served for performance evaluation on unseen data. This decision was made to balance computational efficiency with evaluation reliability, especially considering the high training cost of deep learning models. We used Python (v3.8+) and the Keras/TensorFlow approach for deep learning to train all of the models. All of the data were randomly shuffled before being split, using the same seed value for each shuffle. This made sure that the comparisons were fair and could be repeated. We trained each model using the best hyperparameters and regularization methods. The final evaluation looked at accuracy, precision, recall, and F1-score measures.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003e8.2 Experimental Results\u003c/h2\u003e\u003cp\u003eTo assess the performance of each classifier on Arabic fake news detection, a set of standard evaluation metrics was used, including Accuracy, Precision, Recall, and F1-Score. These metrics were computed across three classification tasks: Main Category (22 classes), Subcategory (75 classes), and Fake/Real (binary). Each model was evaluated using a fixed train-test split (typically 80/20). Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows a summary of the outcomes of all the models that were compared. Traditional machine learning models like RF and Extra Trees did quite well overall, especially on the Fake/Real classification assignment, where F1-scores were over 94%. Among them, the Extra Trees model achieved the highest consistency across all three classification levels.\u003c/p\u003e\u003cp\u003eOn the deep learning side, the CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM model initially provided a solid baseline, particularly on Main Category classification. Further enhancement using Bi-GRU and fine-tuned CNN layers yielded improved results. The Transformer Encoder-based model, on the other hand, did better than all the others in Subcategory classification, with an F1-score of about 88.2%. This is because its self-attention mechanism captures long-range relationships, which is especially useful for Arabic, which has a lot of different forms. In general, CNN\u0026thinsp;+\u0026thinsp;Bi-GRU was the best model for Fake/Real classification, with an accuracy rate of over 93%. For Subcategory classification, Transformer Encoder did the best, and for Main Category predictions, Extra Trees stayed strong. These findings indicate that while traditional models remain competitive on simpler tasks, deep learning and transformer-based approach provide superior handling of hierarchical, multi-label text classification problems in Arabic.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows how well all of the models did in the three categorisation tasks: Main Category, Sub Category, and Fake/Real. The Extra Trees classifier was the most accurate of all the models and did well on all tests. Random Forest and LightGBM are two examples of traditional ensemble approaches that also stayed powerful and stable. Traditional ensemble methods such as Random Forest and LightGBM also maintained strong and stable results. Deep learning models like CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM and CNN\u0026thinsp;+\u0026thinsp;Bi-GRU didn't fare very well at first, but after some tweaking, they got a lot better, especially at the Fake/Real classification job. Transformer-based models also fared well since users could see how things were connected in a certain situation. This comparison shows how well ensemble-based models can generalize and how strong they are when working with structured Arabic news data, especially when it comes to hierarchical and multilabel classification problems.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ecomparative results of all evaluated models\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"13\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c13\" colnum=\"13\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u003cp\u003eMain Category \u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c9\" namest=\"c6\"\u003e\u003cp\u003eSub Category\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c13\" namest=\"c10\"\u003e\u003cp\u003eFake/Real\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eprecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eprecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c11\"\u003e\u003cp\u003eprecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c12\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c13\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRF\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eExtra\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eTrees\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLight\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eGBM\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCNN+\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eBi-LSTM\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCNN\u0026thinsp;+\u0026thinsp;\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eBi-GRU\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTransformer\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.86\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTuned\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCNN\u0026thinsp;+\u0026thinsp;Bi-LSTM\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.79\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.79\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTuned\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCNN\u0026thinsp;+\u0026thinsp;Bi-GRU\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec23\" class=\"Section2\"\u003e\u003ch2\u003e8.3 Discussion\u003c/h2\u003e\u003cp\u003eIn this study, we performed a comparative analysis of several ML and DL models across three key classification tasks: Main Category, Sub Category, and Fake/Real detection. All models were learned using the same data splits, feature normalization, and balanced class distributions to ensure a fair evaluation.\u003c/p\u003e\u003cp\u003eThe Extra Trees classifier did the best overall job of all the ML models. It always got high accuracy and F1 ratings on all jobs, but it was especially good at Sub Category categorisation, which had 75 different classes. The model was a good and useful baseline since it could handle biassed information without needing further oversampling; it trained quickly, and it was able to generalise. Similarly, Random Forest and LightGBM also demonstrated strong and stable performance, though they slightly lagged behind Extra Trees in terms of F1-score, especially for the more granular subcategory task. Their ensemble-based nature allowed them to manage complex, high-dimensional feature spaces effectively. The Transformer-based model did better than all the other deep learning models. Because it might understand the context and pay attention to itself, it could capture the hidden dependencies in Arabic language. This is especially useful for hierarchical and multi-label classification tasks. The Transformer model achieved the highest F1-score in the subcategory and fake/real classification, despite requiring significantly more computational resources.\u003c/p\u003e\u003cp\u003eTuned CNN-based models, such as Tuned CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM and Tuned CNN\u0026thinsp;+\u0026thinsp;Bi-GRU, worked better than their untuned variants, especially when it came to detecting fake and real post. They didn't do better than the Transformer model, but they provided a good balance between performance and training speed. In summary, the results confirm that Extra Trees is the most effective traditional ML model for this task, offering both speed and accuracy. Meanwhile, the Transformer-based model proved to be the most powerful DL architecture, particularly suited for the semantic and structural complexity of Arabic fake news classification.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\u003ch2\u003e8.4 Limitations and Threats to Validity\u003c/h2\u003e\u003cp\u003eAlthough the results of this study appear positive, there are certain problems and potential risks to its validity that should be acknowledged:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eThe dataset comes only from verified Jordanian media sources on Facebook. This restriction makes the results more accurate, but it might be more difficult to use them in other Arabic-speaking areas, on sites like Twitter, or with informal material sources.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eEnsemble models, such as Extra Trees, and deep models like transformers normally work well but can cause overfitting when used on more variable or noisy datasets. Transformers, in particular, require significant amounts of memory and training time, which may not be feasible in real-time or resource-poor environments.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eThe classification assignments have several degrees of hierarchy (for example, 75 subcategories), which makes learning harder. Some of the wrong classifications may have happened because the meanings of closely related categories were too similar.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eWe only tested the models on data from the same domain. To make sure they can be used in other situations, they should test their performance on outside datasets or in cross-lingual contexts in the future.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec25\" class=\"Section2\"\u003e\u003ch2\u003e8.5 Ethical and Legal Considerations in Arabic Fake News Detection\u003c/h2\u003e\u003cp\u003eFinding fake news in Arabic media is hard because of technological and language issues, but it also raises big ethical and legal issues. Laws about false information are often unclear in many Arab countries, and they are sometimes used to restrict free speech. When looking for political or religious content, automated detection systems need to be carefully constructed so that they don't make this kind of excessive authority more serious. Also, protecting users' rights and preventing the possible exploitation of detection technology requires making sure that data gathering and model predictions are confidential, truthful, and accessible.\u003c/p\u003e\u003c/div\u003e"},{"header":"9. CONCLUSION AND FUTURE WORK","content":"\u003cp\u003eThis study presented a comprehensive evaluation of ML and DL models for Arabic fake news detection using a multi-level, multi-label classification approach. The JoNewsFake dataset, developed from authentic Arabic news sources, served as the foundation for training and testing a range of models across three classification tasks: Main Category, Subcategory, and Fake/Real. Experimental results demonstrated that model performance is significantly influenced by the complexity of the task and the richness of the feature representation. Among ensemble machine learning models, the Extra Trees classifier achieved the highest performance, reaching F1 scores of 95% in the main category, 98% in the subcategory, and 95% in the fake/real classification. On the deep learning side, the Transformer-based model achieved the best overall performance, providing a strong balance across all tasks. It demonstrated superior handling of contextual and hierarchical features, with an F1-score of 88.2% in subcategory classification, approximately 85% in main category, and 94% in fake/real classification. The CNN\u0026thinsp;+\u0026thinsp;Bi-GRU model had the highest accuracy for fake/real classification (93%), but it didn't do as well on the other tasks as the Transformer. These results show how well Transformer models can understand the complicated meanings and structures of Arabic news content. This is the first study to compare ensemble machine learning models, Bi-RNN-based deep learning models, and Transformer-based approache for multi-level, multi-label Arabic fake news detection using real news agency data.\u003c/p\u003e\u003cp\u003eIn general, the results show how important it is to use scalable and context-aware models to uncover fake news in Arabic. Extra Trees and other traditional models provide good baselines, but transformer-based designs have a lot of promise to improve multi-level Arabic NLP jobs. The findings underscore the effectiveness of ensemble ML models and transformer-based approach in handling complex, imbalanced, and hierarchical Arabic datasets. The integration of semantic (AraBERT), syntactic (PoS), and emotional features also contributed to performance improvements, especially in challenging subcategory classifications.\u003c/p\u003e\u003cp\u003eFor future research, we could extend this work in several directions. First, adding images, videos, and metadata to multimodal data could make it easier to find false information, especially on social media. Second, using domain-specific data to fine-tune huge pre-trained Arabic transformer models (like MARBERT and QARiB) may help them generalise better and understand context better. Third, using explainable AI (XAI) methods would assist in understanding model predictions, thereby making the systems more transparent and reliable. Also, looking into cross-lingual transfer learning could help Arabic fake news detection models use a lot of resources that are available in English and other languages. Lastly, it would be important for real-world uses to test these models in real-time settings and see how well they hold up with practical scenarios.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cb\u003eFunding Declaration\u003c/b\u003e:\u003c/p\u003e\u003cp\u003eThis research received no external funding.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThe authors confirms sole responsibility for all aspects of the study, including conceptualization, methodology, data collection, analysis, software implementation, visualization, writing, and final approval of the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e\u003cp\u003eThis dataset and accompanying study are part of the first author's PhD thesis at Universiti Malaya.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and analysed during the current study are not publicly available in full because they are under review for separate publication. However, a representative sample has been provided as supplementary material, and the complete dataset is available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAl-Ahmad, B., Al-Zoubi, A. M., Khurma, R. A., \u0026amp; Aljarah, I. (2021). An Evolutionary Fake News Detection Method for COVID-19 Pandemic Information. \u003cem\u003eSymmetry, 13(6), 1091\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eALAZAB, M., AWAJAN, A., ALAZAB, A., KHREISAT, A., ALHYARI, A., \u0026amp; SAADEH, R. (2022). FAKE-NEWS DETECTION SYSTEM USING MACHINE LEARNING ALGORITHMS FOR ARABIC-LANGUAGE. \u003cem\u003eJournal of Theoretical and Applied Information Technology\u003c/em\u003e, 5056-5069.\u003c/li\u003e\n\u003cli\u003eAl-Jalabneh, A. A., \u0026amp; Safori, A. O. (2022). \u003cem\u003eCovid-19 and Misinformation Prevalence: A Content Analysis of Fake News Stories Spread in Jordan.\u003c/em\u003e In The Implementation of Smart Technologies for Business Success and Sustainability: During COVID-19 Crises in Developing Countries (pp. 535-545). Cham: Springer International Publishing.\u003c/li\u003e\n\u003cli\u003eAljedani, N., Alotaibi, R., \u0026amp; Taileb, M. (2020). Multi-Label Arabic Text Classification: An Overview. \u003cem\u003eInternational Journal of Advanced Computer Science and Applications, 11(10).\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eAljohani, E. (2024). Enhancing Arabic Fake News Detection: Evaluating Data Balancing Techniques Across Multiple Machine Learning Models. \u003cem\u003eEngineering, Technology \u0026amp; Applied Science Research, \u003c/em\u003e, 14(4), 15947-15956.\u003c/li\u003e\n\u003cli\u003eAlkhair, M., Meftouh, K., Othman, N., \u0026amp; Sma\u0026iuml;li, K. (2019). An arabic corpus of fake news: Collection, analysis and classification.\u0026quot; . \u003cem\u003eArabic Language Processing: From Theory to Practice: 7th International Conference, ICALP 2019, Nancy, France, October 16\u0026ndash;17, 2019, Proceedings 7. Springer International Publishing\u003c/em\u003e, pp. 292-302.\u003c/li\u003e\n\u003cli\u003eAlmarashy, A. H., Feizi-Derakhshi, M.-R., \u0026amp; Salehpour, P. (2023). Enhancing fake news detection by multi-feature classification. \u003cem\u003eIEEE Access, 11, 139601-139613.\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eAlmuzaini, H. A., \u0026amp; Azmia, A. M. (2022). An unsupervised annotation of Arabic texts using multi-label topic modeling and genetic algorithm. \u003cem\u003eExpert Systems with Applications, 203, 117384.\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eAlruily, M. (2021). Classification of Arabic Tweets: A Review. \u003cem\u003eElectronics, 10(10), 1143\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eAl-Salemia, B., Ayoba, M., Kendallb, G., \u0026amp; Noaha, S. A. (2019). \u003cem\u003eMulti-label Arabic text categorization: A benchmark and baseline comparison of multi-label learning algorithms.\u003c/em\u003e Information Processing \u0026amp; Management, 56(1), 212-227.\u003c/li\u003e\n\u003cli\u003eAl-Taie, M. Z. (2025). Comparative Study of Machine Learning Approaches for Detecting Fake News in Arabic Text. \u003cem\u003eIETI Transactions on Data Analysis and Forecasting, 3(1), \u003c/em\u003e, 18-31.\u003c/li\u003e\n\u003cli\u003eAlturayeif, N., Luqman, H., \u0026amp; Ahme, M. (2022). MAWQIF: A Multi-label Arabic Dataset for Target-specific Stance Detection. \u003cem\u003eIn Proceedings of the Seventh Arabic Natural Language Processing Workshop (WANLP)\u003c/em\u003e, 174-184.\u003c/li\u003e\n\u003cli\u003eAl-Yahya, M., Al-Khalifa, H., Al-Baity, H., AlSaeed, D., \u0026amp; Essam, A. (2021). Arabic Fake News Detection: Comparative Study of Neural Networks and Transformer-Based Approaches. \u003cem\u003eComplexity, vol. 2021\u003c/em\u003e, Article ID 5516945, 10 pages.\u003c/li\u003e\n\u003cli\u003eAlzahrani, L., \u0026amp; Al-yahya, M. (2024). Pre-Trained Language Model Ensemble for Arabic Fake News Detection. \u003cem\u003eMathematics, 12(18), \u003c/em\u003e, 1-17.\u003c/li\u003e\n\u003cli\u003eAmeur, M. S., \u0026amp; Aliane, H. (2021). \u003cem\u003eARACOVID19-MFH: ARABIC COVID-19 MULTI-LABEL FAKE NEWS AND HATE SPEECH DETECTION DATASET.\u003c/em\u003e Procedia Computer Science, 189, 232-241.\u003c/li\u003e\n\u003cli\u003eBarua, R., Rahman, M. M., \u0026amp; Joy, U. G. (2025). Comparative analysis of Bangla news classification: a study of fake news detection and multiclass classification using BERT and FastText. \u003cem\u003eInternational Journal of Computers and Applications\u003c/em\u003e, 47(5), 475-485.\u003c/li\u003e\n\u003cli\u003eBelinda, M. J., David, S. A., Kannan, E., Naveena, N. R., Rajathi, K., \u0026amp; Mahesh, C. (2024, October). Fake News Detection Using Logistic Regression and LightGBM. \u003cem\u003eIn 2024 2nd International Conference on Self Sustainable Artificial Intelligence Systems (ICSSAS)\u003c/em\u003e, (pp. (pp. 1-5). IEEE.).\u003c/li\u003e\n\u003cli\u003eBsoul, M. A., Qusef, A., \u0026amp; Abu-Soud, S. (2022). Building an Optimal Dataset for Arabic Fake News Detection. \u003cem\u003eProcedia Computer Science, 201\u003c/em\u003e, 665-672.\u003c/li\u003e\n\u003cli\u003eDhiman, P., Kaur, A., Gupta, D., Juneja, S., Nauman, A., \u0026amp; Muhammad, G. (2024). GBERT: A hybrid deep learning model based on GPT-BERT for fake news detection. \u003cem\u003eHeliyon, 10(16).\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eFauzan, F. J., Putri, C. M., \u0026amp; Laura, P. (2024). Performance Comparison of Classification Algorithms for Chronic Kidney Disease Prediction. \u003cem\u003eIJATIS: Indonesian Journal of Applied Technology and Innovation Science\u003c/em\u003e, 1(2), 68-74.\u003c/li\u003e\n\u003cli\u003eFouad, K. M., Sabbeh, S. F., \u0026amp; Medhat, W. (2022). Arabic Fake News Detection Using Deep Learning. \u003cem\u003eComputers, Materials \u0026amp; Continua, 71(2).\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eHadeel, S., Mohamed, E., \u0026amp; Orasan, C. (2020). Fake or Real? A Study of Arabic Satirical Fake News. \u003cem\u003earXiv preprint arXiv:2011.00452\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eImani, M., Beikmohammadi, A., \u0026amp; Arabnia, H. R. (2025). Comprehensive Analysis of Random Forest and XGBoost Performance with SMOTE, ADASYN, and GNUS Under Varying Imbalance Levels. \u003cem\u003eTechnologies, 13(3), \u003c/em\u003e, 88.\u003c/li\u003e\n\u003cli\u003eImran, M., Aslam, N., Ahmad, H., Mazhar, F., Bhatti, Y. I., \u0026amp; Abid, U. (2024). Predictive modeling of chronic kidney disease using extra tree classifier: A comparative analysis with traditional methods. \u003cem\u003eJournal of Computing \u0026amp; Biomedical Informatics\u003c/em\u003e, 6(02), 261-271.\u003c/li\u003e\n\u003cli\u003eJain, M. K., Gopalani, D., \u0026amp; Meena, Y. K. (2025). Hybrid CNN-Bi-LSTM model with HHO feature selection for enhanced fake news detection. \u003cem\u003eSocial Network Analysis and Mining, 15(1), 43.\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eKaur, S., Kumar, P., \u0026amp; Kumaraguru, P. (2020). Automating fake news detection system using multi-level voting model. \u003cem\u003eSoft Computing, 24(12), 9049-9069.\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003eKhalil, A., Jarrah, M., \u0026amp; Aldwairi, M. (2024). Hybrid Neural Network Models for Detecting Fake News Articles. \u003cem\u003eHuman-Centric Intelligent Systems, 4(1)\u003c/em\u003e, 136-146.\u003c/li\u003e\n\u003cli\u003eKhalil, A., Jarrah, M., Aldwairi, M., \u0026amp; Jaradat, M. (2022). AFND: Arabic fake news dataset for the detection and classification of articles credibility. \u003cem\u003eData in Brief\u003c/em\u003e. doi:https://doi.org/10.1016/j.dib.2022.108141\u003c/li\u003e\n\u003cli\u003eLiu, Y. M., Zhao, L., Liang, Y., Zhang, P., \u0026amp; Jin, B. (2022). Hybrid embedding-based text representation for hierarchical multi-label text classification. \u003cem\u003eExpert Systems with Applications, \u003c/em\u003e, 187, 115905.\u003c/li\u003e\n\u003cli\u003eMaxwell, A., Li, R., B., Y., Weng, H., Ou, A., Hong, H., \u0026amp; ... \u0026amp; Zhang, C. (2017). Deep learning architectures for multi-label classification of intelligent health risk prediction. \u003cem\u003eBMC bioinformatics, 18\u003c/em\u003e, 121-131.\u003c/li\u003e\n\u003cli\u003eNARRA, M., MUHAMMADUMER, SADIQ, S., ESHMAWI, A., KARAMTI, H., MOHAMED, A., \u0026amp; ASHRAF, I. (2022). Selective Feature Sets Based Fake News Detection for COVID-19 to Manage Infodemic. \u003cem\u003eIEEE Access, 10, \u003c/em\u003e, 98724-98736.\u003c/li\u003e\n\u003cli\u003eNassif, A. B., Elnagar, A., Elgendy, O., \u0026amp; Afadar, Y. (2022). Arabic fake news detection based on deep contextualized embedding models. \u003cem\u003eNeural Computing and Applications, 34(18), \u003c/em\u003e, 16019-16032.\u003c/li\u003e\n\u003cli\u003ePatil, D. R. (2022). Fake News Detection Using Majority Voting Technique . \u003cem\u003earXiv preprint arXiv:2203.09936.\u003c/em\u003e \u003c/li\u003e\n\u003cli\u003ePillai, A. S. (2024). FAKE NEWS MULTI CLASS DETECTION USING TRANSFORMERS AND GRADIENT BOOSTING ENSEMBLE. \u003cem\u003eInternational Research Journal of Modernization in Engineering Technology and Science\u003c/em\u003e, https://www. researchgate. net/publication/378691966_FAKE_NEWS_MULTI_CLAS S_DETECTION_USING_TRANSFORMERS_AND_GRADIENT_BOOSTING_EN SEMBLE.\u003c/li\u003e\n\u003cli\u003eQandos, N., Hamad, G., Alharb, M., Alturki, S., Alharbi, W., \u0026amp; Albelaihi, A. A. (2024). Multiscale cascaded domain-based approach for Arabic fake reviews detection in e-commerce platforms. \u003cem\u003eJournal of King Saud University-Computer and Information Sciences, 36(2), 101926\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eRahmanian, E. (2023). Fake news: a classification proposal and a future research agenda. \u003cem\u003eSpanish Journal of Marketing-ESIC, 27(1),\u003c/em\u003e, 60-78.\u003c/li\u003e\n\u003cli\u003eSalem, F. K., Feel, R. A., Elbassuoni, S., Jaber, M., \u0026amp; Farah, M. (2019, July). \u003cem\u003eFa-kes: A fake news dataset around the syrian war.\u003c/em\u003e In Proceedings of the international AAAI conference on web and social media (Vol. 13, pp. 573-582).\u003c/li\u003e\n\u003cli\u003eSOROUR, S. E., \u0026amp; ABDELKADER, H. E. (2022). AFND: ARABIC FAKE NEWS DETECTION WITH AN ENSEMBLE DEEP CNN-LSTM MODEL. \u003cem\u003eJ. Theor. Appl. Inf. Technol, 100(14)\u003c/em\u003e, 5072-5086.\u003c/li\u003e\n\u003cli\u003eTarekegn, A. N., Ullah, M., \u0026amp; Cheikh, F. A. (2024). Deep learning for multi-label learning: A comprehensive survey. \u003cem\u003earXiv preprint arXiv\u003c/em\u003e, 2401.16549.\u003c/li\u003e\n\u003cli\u003eThaher, T., Saheb, M., Turabieh, H., \u0026amp; Chantar, H. (2021). Intelligent Detection of False Information in Arabic Tweets Utilizing Hybrid Harris Hawks Based Feature Selection and Machine Learning Models. Symmetry, 13(4), 556.\u003c/li\u003e\n\u003cli\u003eWotaifi, T. A., \u0026amp; Dhannoon, B. N. (2023). An Effective Hybrid Deep Neural Network for Arabic Fake News Detection. \u003cem\u003eBaghdad Science Journal, 20(4)\u003c/em\u003e, 1392-1392.\u003c/li\u003e\n\u003cli\u003eYildirim, O., Bakhshi, S., \u0026amp; Can, F. (2024, October). Prioritized Binary Transformation Method for Efficient Multi-label Classification of Data Streams with Many Labels. \u003cem\u003eIn Proceedings of the 33rd ACM International Conference on Information and Knowledge Management \u003c/em\u003e, (pp. (pp. 4218-4222)).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://newsinitiative.withgoogle.com/\u003c/span\u003e\u003cspan address=\"https://newsinitiative.withgoogle.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Arabic fake news detection, multi-level classification, JoNewsFake dataset, transformer model, deep learning, ensemble learning, AraBERT embeddings, Arabic NLP","lastPublishedDoi":"10.21203/rs.3.rs-7424709/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7424709/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe public's trust and the integrity of information are at risk by the spread of fake news on Arabic social media platforms. Arabic is still underrepresented in multi-label multi-level fake news detection because of its linguistic complexity and lack of resources, despite significant efforts being made in English. This study provides a full comparison of machine learning and deep learning models for multi-label, multi-level Arabic fake news classification. Using the newly constructed JoNewsFake dataset collected from verified Jordanian news agencies, the models were trained to classify news into main categories, subcategories, and fake/real labels. Experiments included traditional ML classifiers (Random Forest, Extra Trees, and LightGBM) and advanced DL models (CNN\u0026thinsp;+\u0026thinsp;Bi-LSTM, CNN\u0026thinsp;+\u0026thinsp;Bi-GRU, fine-tuned CNN variants, and Transformer-based). Results showed that the Extra Trees classifier outperformed all ML and DL models, achieving F1-scores of 95% (Main Category), 98% (Subcategory), and 95% (Fake/Real). Among the DL models, the transformer-based model yielded the best performance for subcategory classification F1 at 88.2%, while CNN\u0026thinsp;+\u0026thinsp;Bi-GRU achieved 93% accuracy in binary fake/real classification. The study confirms the value of combining syntactic, semantic, and emotional features (Tuned AraBERT) to boost classification performance, particularly for complex subcategory tasks. These findings lay the groundwork for future enhancements in Arabic fake news detection using explainable and scalable models.\u003c/p\u003e","manuscriptTitle":"An Empirical Comparison of Ensemble model and Deep Learning Models for Multi-Level Arabic Fake News Classification using JoNewsFake Dataset","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-16 16:56:19","doi":"10.21203/rs.3.rs-7424709/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8611ebce-a508-4153-946e-2534716f5c73","owner":[],"postedDate":"September 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":54687759,"name":"Physical sciences/Engineering"},{"id":54687760,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-02-17T19:34:08+00:00","versionOfRecord":{"articleIdentity":"rs-7424709","link":"https://doi.org/10.7717/peerj-cs.3510","journal":{"identity":"peerj-computer-science","isVorOnly":true,"title":"PeerJ Computer Science"},"publishedOn":"2026-02-04 00:00:00","publishedOnDateReadable":"February 4th, 2026"},"versionCreatedAt":"2025-09-16 16:56:19","video":"","vorDoi":"10.7717/peerj-cs.3510","vorDoiUrl":"https://doi.org/10.7717/peerj-cs.3510","workflowStages":[]},"version":"v1","identity":"rs-7424709","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7424709","identity":"rs-7424709","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}