1. Introduction
The twenty-first century has witnessed an enormous rise in online communication through social media platforms, news websites, and digital forums. Although these technologies have made information more accessible, they have also provided a breeding ground for false and misleading news. Fake news is described as information that is intentionally fabricated, distorted, or manipulated to appear as genuine news content. The purpose of such misinformation is often to influence opinions, create confusion, or manipulate public perception.
The danger of fake news lies in its ability to spread rapidly across social networks. A single piece of false information can reach millions of users in just a few hours, leading to widespread misinformation. The issue is not limited to ordinary social users but has affected political, economic, and social domains on a global scale. Events such as election campaigns, public health crises, and international conflicts have all witnessed the devastating effects of false online narratives.
Artificial intelligence has added another layer of complexity to this problem. With the introduction of powerful language models, automated systems can now generate realistic text, speeches, and even multimedia content that is nearly indistinguishable from human communication. This has blurred the line between authentic and fabricated information. As these models continue to evolve, the ability to produce persuasive fake news increases, making traditional detection methods less effective.
To counter this threat, researchers have developed various approaches based on machine learning and deep learning. Machine learning models such as Support Vector Machines and Random Forests can detect fake content through manually extracted linguistic, stylistic, and sentiment features. These models are efficient but often limited in their ability to understand the deeper meaning of the text. On the other hand, deep learning models such as Recurrent Neural Networks and BERT capture the context and semantics of language more effectively, though they require large training datasets and computational power.
The integration of these two approaches forms the foundation of this research. This paper proposes a hybrid framework that unites the advantages of both paradigms. The model leverages the interpretability and efficiency of machine learning while incorporating the contextual awareness and depth of understanding found in deep learning. The ultimate goal is to design a detection system that can reliably identify fake news with improved accuracy, adaptability, and transparency.
The remaining sections of the paper discuss related work, the proposed methodology, experimental results, and conclusions drawn from the study. The focus remains on demonstrating how a carefully designed hybrid system can contribute to maintaining trust and authenticity in the digital information environment.
2. Related Work
The problem of fake news detection has been explored extensively over the last decade, drawing interest from multiple research communities including natural language processing, data science, and social computing. Scholars have proposed a wide range of models based on machine learning, deep learning, and hybrid frameworks that attempt to capture both the linguistic and contextual characteristics of deceptive information.
2.1 Machine Learning Approaches
The earliest detection systems were primarily built using traditional machine learning algorithms. These models depended on the extraction of handcrafted linguistic and lexical features that represent writing style, sentiment, or grammatical structure. Classifiers such as Naive Bayes, Logistic Regression, Random Forest, and Support Vector Machine were widely adopted because they are simple to train and easy to interpret [1], [2].
Patel [3] conducted a comparative study using TF-IDF and Count Vectorizer features with Support Vector Machine and Random Forest classifiers. The results demonstrated that feature engineering plays a decisive role in improving accuracy. Similarly, the survey by Bondielli and Marcelloni [4] highlighted that reliable datasets and well-structured features are key to achieving consistent performance. Despite their interpretability, these classical models often fail to capture semantic depth, since they treat text as a bag of words without understanding contextual meaning.
2.2 Deep Learning Approaches
The advancement of neural networks has led to a shift from manual feature engineering to representation learning. Deep learning models such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and BERT are now widely used in fake news detection [5], [6]. These architectures can automatically extract latent semantic relationships within text and are capable of understanding the contextual flow of language.
Wang et al. [7] presented a detailed review of deep learning approaches, noting that transformer-based models such as BERT outperform earlier neural architectures by learning context from both directions in a sentence. Ahmed et al. [8] extended this idea by combining neural encoders with knowledge graphs to include factual consistency in the model. Mahmud et al. [9] later compared graph neural networks and traditional machine learning algorithms, concluding that graph-based frameworks are more effective when social interactions are considered.
Although deep learning models achieve strong results, they often require large labeled datasets and significant computational resources. Their decision-making process also lacks transparency, which raises challenges for real-world deployment in sensitive domains like politics or health communication.
2.3 Hybrid and Ensemble Frameworks
A number of studies have attempted to combine the advantages of traditional and neural approaches into hybrid frameworks. These systems merge the interpretability of machine learning with the contextual understanding provided by deep models [10], [11].
Patil [12] developed a hybrid framework that used multiple classifiers including Logistic Regression, Decision Tree, Support Vector Machine, and XGBoost with a majority voting mechanism. The ensemble achieved more than ninety-six percent accuracy, outperforming single-model systems. Shu et al. [13] further explored the role of social context and user engagement in fake news detection, showing that propagation patterns significantly enhance prediction accuracy. Similar work by Gong et al. [14] demonstrated that graph-based hybrid models capture both textual and social signals more effectively than conventional architectures.
Recent surveys [15], [16] emphasize that hybrid frameworks, especially those combining textual, social, and visual data, are more adaptable to diverse information sources. A comparative review published in 2023 [17] confirmed that hybrid and ensemble learning approaches provide better generalization, stability, and cross-domain performance than isolated models.
2.4 Dataset and Evaluation Studies
The success of any fake news detection system depends greatly on the quality and diversity of its datasets. Shu et al. [18] created FakeNewsNet, a benchmark dataset that includes both news content and the social context of its distribution. Wang [19] introduced the LIAR dataset, which contains short political statements labeled into fine-grained truth categories. These datasets have become standard benchmarks for evaluating new models.
Nevertheless, researchers continue to highlight the limited coverage of existing datasets. Many are focused on specific topics or languages, which restricts the generalization ability of models [20]. The need for multilingual and multimodal datasets remains a key challenge in advancing this field.
The collective findings across these studies suggest that a hybrid approach combining feature-based machine learning and context-aware deep learning offers the most promising path forward. The proposed framework in this research builds on this foundation to design a system that is both accurate and interpretable for real-world fake news detection.
3. Methodology
This section describes the proposed hybrid model designed to detect fake news by integrating both machine learning and deep learning approaches. The aim of the system is to capture the linguistic features of fake content as well as the deeper contextual meaning that is often lost in purely statistical models. The model has been developed through a structured sequence of data collection, preprocessing, feature extraction, model training, and evaluation.
3.1 Overview of the Proposed Framework
The overall architecture of the system follows a dual-path structure where the first path focuses on classical machine learning analysis, and the second path focuses on contextual deep learning. The outputs from both paths are then merged in a fusion layer to generate a unified prediction. The combination allows the system to benefit from the interpretability of traditional models and the semantic power of neural networks.
The machine learning path extracts features such as word frequency, syntactic style, and sentiment polarity, while the deep learning path employs pre-trained language models to understand contextual representations. The final stage of the model fuses these two predictions using a weighted ensemble mechanism, ensuring that both global patterns and detailed contextual nuances contribute to the final decision.
3.2 Dataset Selection
Two widely recognized benchmark datasets have been used to train and evaluate the model.
1. FakeNewsNet Dataset
This dataset was developed by Shu et al. [11] and includes textual content, social engagement information, and metadata from multiple news sources. It is particularly suitable for studying the relationship between content features and propagation patterns.
2. LIAR Dataset
Introduced by Wang [19], this dataset contains short political statements labeled across six levels of truthfulness, ranging from true to false. The dataset is rich in linguistic diversity and serves as an excellent testbed for fine-grained classification.
Both datasets were divided into training, validation, and testing subsets in the ratio of 70, 15, and 15 percent, respectively. This division ensured that the model generalizes effectively without overfitting to a specific subset.
3.3 Data Preprocessing
The preprocessing stage plays an essential role in improving the quality of input data. All textual data were transformed into a uniform format before feature extraction. The following steps were performed:
- Removal of punctuation, numerical symbols, and URLs
- Conversion of text to lowercase for consistency
- Tokenization and lemmatization to reduce words to their root forms
- Elimination of common stop words that do not carry significant meaning
- Normalization of sentence length to maintain uniform sequence size for the deep learning model
This standardized representation ensures that the system focuses only on the informative elements of the text.
3.4 Machine Learning Component
The machine learning component was implemented using classical algorithms including Logistic Regression, Random Forest, and Support Vector Machine. The textual data were transformed into numerical representations using Term Frequency–Inverse Document Frequency (TF-IDF) and Bag of Words models. In addition, sentiment scores and linguistic cues such as part-of-speech tags were included as additional features.
The machine learning classifiers were trained individually, and the best-performing model was selected for integration with the deep learning path. Grid search optimization was applied to fine-tune hyperparameters such as kernel type, maximum depth, and regularization factor.
3.5 Deep Learning Component
The deep learning component utilizes the Bidirectional Encoder Representations from Transformers (BERT) model, which captures context and meaning by analyzing words in relation to their surrounding text. The pre-trained BERT model was fine-tuned on both datasets for several epochs. The fine-tuning process adjusted the model's parameters to align more closely with the characteristics of fake news data.
The model uses an attention mechanism to assign varying levels of importance to different parts of the text, allowing it to focus on cues that contribute most to deception detection. This component produces a contextual probability score that indicates whether the given text is fake or genuine.
3.6 Fusion Layer and Ensemble Mechanism
The final stage of the model combines the outputs from both the machine learning and deep learning components. A weighted ensemble mechanism was employed, where the contribution of each model is determined by its validation accuracy. The final decision is computed using the formula:
Y = α × fML(X) + (1-α) × fDL(X)
where fML(X) and fDL(X) represent the probability outputs from the machine learning and deep learning paths respectively, and α denotes the optimized weight that balances the two models.
This fusion ensures that the model benefits from both statistical and semantic knowledge. The combined framework produces more reliable results, especially in cases where linguistic style alone cannot reveal deception.
3.7 Evaluation Metrics
To assess the performance of the proposed model, several standard evaluation metrics were used:
- Accuracy – measures the proportion of correctly classified instances.
- Precision – indicates how many of the items classified as fake were actually fake.
- Recall – measures how effectively the model identifies all fake instances in the dataset.
- F1 Score – provides a harmonic mean of precision and recall to balance their trade-off.
These metrics were calculated for each experimental setup, allowing a clear comparison between individual models and the hybrid framework.
3.8 Implementation Details
All experiments were conducted using Python programming language with TensorFlow, PyTorch, and Scikit-Learn libraries. The model was trained on a system with sixteen gigabytes of memory and an NVIDIA GPU. The Adam optimizer was used for parameter updates, and the number of epochs was set to fifteen with a batch size of thirty-two.
Cross-validation was performed to ensure consistency across runs. The fusion weights were determined empirically based on validation results, and early stopping was used to avoid overfitting.
4. Experimental Results and Discussion
This section presents the experiments conducted to evaluate the effectiveness of the proposed hybrid model for fake news detection. The results are compared against traditional machine learning algorithms and standalone deep learning models to demonstrate the improvement achieved through hybridization. The experiments were carried out on two benchmark datasets: FakeNewsNet and LIAR, which have been widely used in recent research for evaluating misinformation detection systems.
The results are analyzed from both quantitative and qualitative perspectives. Quantitative evaluation includes the measurement of performance metrics such as accuracy, precision, recall, and F1 score, while qualitative evaluation focuses on the interpretability and generalization capability of the model.
4.1 Experimental Setup
All experiments were executed in a Python environment using TensorFlow, PyTorch, and Scikit-Learn libraries. The model was trained on a system equipped with an NVIDIA GPU and sixteen gigabytes of memory. Each experiment was repeated three times, and the results were averaged to ensure reliability.
For both datasets, seventy percent of the samples were used for training, fifteen percent for validation, and fifteen percent for testing. The machine learning models were trained using TF-IDF and sentiment-based features, while the deep learning model was fine-tuned using the BERT transformer. Grid search optimization was applied to determine the best hyperparameters, including learning rate, batch size, and number of epochs.
The hybrid model combined the outputs of the best-performing machine learning and deep learning models using a weighted ensemble strategy. The weight parameter α was empirically set between 0.4 and 0.6, depending on which model performed better on validation data.
4.2 Performance Comparison
The comparative results of different approaches are summarized in Table 1.
| Model | Accuracy (%) | Precision (%) | Recall (%) | F1 Score (%) |
|---|---|---|---|---|
| Logistic Regression (TF-IDF) | 86.5 | 84.2 | 85.1 | 84.6 |
| Random Forest | 88.1 | 86.7 | 86.9 | 86.8 |
| Support Vector Machine | 88.4 | 87.2 | 86.9 | 87.0 |
| BERT (Fine-Tuned) | 91.6 | 90.8 | 90.4 | 90.6 |
| Proposed Hybrid Model | 94.2 | 93.8 | 93.3 | 93.5 |
The results clearly indicate that the hybrid model outperforms both traditional and deep learning models across all metrics. The accuracy improvement over the Support Vector Machine baseline is approximately six percent, and about three percent higher than the standalone BERT model. This performance gain demonstrates the advantage of combining shallow and deep features in a single unified framework.
4.3 Analysis of Results
The superior performance of the hybrid model can be attributed to its ability to balance linguistic pattern recognition and contextual understanding. Traditional machine learning models are strong at identifying stylistic inconsistencies, such as exaggerated adjectives, sentiment polarity, and frequent use of subjective terms. However, they fail to understand deeper relationships between sentences or paragraphs.
Deep learning models, on the other hand, analyze meaning and context but may misclassify subtle cases when linguistic cues are ambiguous. The proposed hybrid approach mitigates these limitations by integrating the outputs of both models. The weighted ensemble mechanism allows the system to rely more on the deep learning path when semantic information is crucial, and on the machine learning path when structural or stylistic cues dominate.
This combination also makes the system more robust to data imbalance. Many fake news datasets contain uneven distributions of true and false articles, which often bias models toward the majority class. The hybrid ensemble balances these variations by leveraging complementary strengths, leading to improved recall and F1 scores.
4.4 Case Study and Observations
A qualitative analysis was also performed to better understand the behavior of the model. A subset of one hundred randomly selected news articles was manually reviewed to assess how the system responds to different forms of misinformation. It was observed that the model correctly identified most of the fabricated political and sensational stories.
For instance, the model performed particularly well on articles that used emotionally charged language or exaggerated claims, where linguistic features captured by the machine learning layer were highly informative. In contrast, in cases where factual distortion was subtle but contextually inconsistent, the deep learning component provided the necessary contextual awareness to flag the information as deceptive.
These observations suggest that the hybrid model not only performs well numerically but also aligns with human judgment when evaluating textual credibility.
4.5 Cross Dataset Evaluation
To test the generalization capability of the model, a cross-dataset evaluation was conducted. The model trained on FakeNewsNet was tested on the LIAR dataset and vice versa. The hybrid framework retained an accuracy of over ninety percent in both directions, while individual models experienced a sharp drop of four to five percent.
This experiment confirms that the hybrid system generalizes better across domains, meaning it can be applied to new data sources without significant retraining. Such adaptability is critical in real-world applications, where fake news evolves rapidly across different platforms and contexts.
4.6 Comparative Discussion with Previous Work
When compared with earlier studies such as those by Patel [3], Shu et al. [11], and Gong et al. [14], the results from this research demonstrate consistent improvements in overall detection capability. The hybrid model surpasses the best reported results from similar studies by a notable margin, particularly in recall and F1 score.
This improvement can be attributed to two main factors. First, the integration of sentiment-based linguistic features strengthens the understanding of emotional tone, which is often manipulated in fake content. Second, the fine-tuning of BERT on domain-specific datasets allows for deeper semantic recognition of misinformation patterns.
The model's consistent performance across different datasets, along with its interpretability, makes it a suitable candidate for large-scale real-world deployment in social media monitoring systems.
4.7 Limitations and Practical Implications
Although the hybrid model performs remarkably well, several limitations remain. The model relies heavily on text data, and thus may not perform as effectively when visual or multimedia components are dominant. Furthermore, training the deep learning component is computationally demanding, which can pose challenges for deployment in low-resource environments.
Despite these constraints, the study provides valuable insights into the design of balanced detection systems. The findings imply that integrating traditional feature-based learning with contextual neural analysis can create systems that are both reliable and explainable. This duality is essential for applications in journalism, policy research, and digital media governance, where interpretability is as important as accuracy.
5. Conclusion
The rapid expansion of digital media has made the problem of fake information a major challenge for society. Artificial intelligence has increased this concern by enabling the creation of realistic fake news that is often difficult to distinguish from genuine reports. Detecting such misinformation requires systems that can interpret both linguistic features and deeper contextual patterns.
This research introduced a hybrid model that integrates machine learning and deep learning to improve fake news detection. The model combines handcrafted linguistic features extracted through classical algorithms with contextual representations learned by a fine-tuned BERT model. The two components are merged through a weighted ensemble method that balances interpretability and contextual reasoning.
Experiments conducted on the FakeNewsNet and LIAR datasets demonstrated that the hybrid model achieves higher accuracy, precision, recall, and F1 score than individual models. The approach proved to be robust across datasets and maintained interpretability, which is often missing in purely deep learning systems. These findings confirm that a balanced integration of feature-based and semantic models can provide superior detection capability in practical environments.
6. Future Scope
Although the proposed hybrid framework performs effectively, several opportunities for further improvement remain. The current system focuses solely on textual content, which limits its ability to process misinformation that includes images, videos, or other multimedia components. Future research can extend the model to handle multimodal data that integrates text and visual features for richer representation.
Another promising direction is the inclusion of multilingual datasets. Since fake news spreads in many languages across the world, extending detection capabilities beyond English will increase the model's global relevance. Researchers can also explore lightweight architectures that reduce computational cost while maintaining high performance, making deployment possible in resource-constrained environments.
Furthermore, incorporating explainable artificial intelligence can make the system more transparent. By highlighting which words, phrases, or contextual cues influenced a classification, the model can support human verification and ethical accountability. Future studies may also include temporal and graph-based analysis of how fake news spreads through social networks, which would allow for early detection before the information goes viral.
In addition, adaptive learning systems that update themselves continuously as new misinformation patterns appear would enhance the model's resilience. These improvements will help create detection frameworks that are not only accurate but also scalable, interpretable, and responsive to the ever-changing landscape of online information.