Meta-Analysis of Feature Representation Techniques for Fake News Detection: Evidence from Publicly Available Benchmark Datasets

Technical Summary: Meta-Analysis of Feature Representation Techniques for Fake News Detection

Problem Statement
The rapid proliferation of digital media and social networking platforms has accelerated the dissemination of fake news, creating significant social, political, and economic challenges. While automated detection using machine learning (ML) is a promising solution, a critical bottleneck remains: the lack of a universally accepted feature representation technique that performs consistently across diverse datasets, languages, and domains. Existing research often relies on single feature extraction methods or isolated classifiers, failing to capture the full spectrum of linguistic characteristics (both lexical importance and contextual relationships) inherent in deceptive content. Furthermore, the variability in benchmark datasets and evaluation protocols makes direct comparison of reported accuracies difficult, creating uncertainty regarding the most efficient and effective detection strategies.

Methodology
This study employs a secondary-data meta-analysis and systematic comparative review rather than proposing a new experimental model. The research synthesizes empirical evidence from previously published studies summarized in a source thesis, focusing on publicly available benchmark datasets (e.g., Bangla Fake News, Twitter Fake News, FakeNewsNet, LIAR, WELFake).

The methodology follows a six-stage workflow:

1. Identification: Selecting peer-reviewed studies utilizing ML-based fake news detection with specific feature representation techniques.
2. Selection: Filtering for studies that focus on fake news, evaluate text feature representations, and report measurable performance on benchmark datasets.
3. Data Extraction: Systematically extracting variables including dataset characteristics, feature extraction methods (TF-IDF, N-gram, BoW, Word2Vec, mBERT, BERT), classifiers, ensemble strategies, and performance metrics (Accuracy, Precision, Recall, F1-score).
4. Comparative Analysis: Organizing findings to identify methodological patterns, strengths, and limitations across different experimental conditions.
5. Evidence Synthesis: Performing a narrative meta-analysis to integrate recurring patterns regarding feature engineering effectiveness.
6. Interpretation: Drawing conclusions on current trends, gaps, and future directions based on the synthesized evidence.

The study treats feature representation techniques as independent variables and reported detection performance as the dependent variable, with dataset characteristics and classifier selection acting as moderating variables.

Key Contributions and Results
The analysis yields several critical findings regarding the interplay between feature engineering and classifier performance:

• Superiority of Hybrid Representations: The study concludes that hybrid feature representations generally outperform individual extraction methods. Specifically, combining TF-IDF (which captures term importance) with N-gram (which preserves local word sequences and contextual patterns) consistently yields higher classification accuracy. This combination allows classifiers to recognize both informative terms and meaningful phrase-level patterns without the substantial computational overhead of deep learning.
• Effectiveness of Traditional ML: Traditional statistical methods, particularly TF-IDF combined with linear classifiers like Support Vector Machines (SVM) and Logistic Regression, remain highly competitive. In several reviewed instances (e.g., Hossen et al.), TF-IDF-based stacking ensembles achieved performance comparable to or exceeding deep learning models (up to 99.6% accuracy) while requiring significantly fewer computational resources.
• Role of Ensemble Learning: Ensemble strategies (Hard Voting, Soft Voting, Stacking) consistently demonstrate greater predictive stability and robustness than individual classifiers across different datasets. Stacking, which utilizes a meta-classifier to combine base learner outputs, often achieves the highest reported performance but introduces higher implementation complexity.
• Trade-offs with Deep Learning: While transformer-based models (BERT, mBERT) and hybrid deep learning architectures (CNN-LSTM-BERT) frequently achieve state-of-the-art accuracy (e.g., up to 99.74%), they demand extensive computational resources, large annotated datasets, and long training times. Their practical deployment is often hindered by these constraints and reduced interpretability compared to traditional ML approaches.
• Dataset Dependency: No single feature representation is universally optimal. Performance is heavily moderated by dataset characteristics, including language, size, class balance, and domain. High accuracy on one dataset does not guarantee generalizability to another.

Significance of the Study
The paper positions its contribution as a consolidation of existing empirical evidence to guide researchers and practitioners toward more efficient and practical detection systems. Its significance lies in:

1. Validating Feature Engineering: It establishes that effective feature engineering is as critical as classifier selection. The study challenges the assumption that transformer-based models are the only viable path, demonstrating that carefully engineered hybrid features (TF-IDF + N-gram) with lightweight ML models offer a computationally efficient and highly effective alternative.
2. Bridging the Gap: It provides a broader perspective on methodological effectiveness by synthesizing findings across multiple benchmark datasets, moving beyond isolated experimental outcomes to identify consistent trends.
3. Practical Guidance: For practitioners and platform developers, the study offers a roadmap for selecting appropriate techniques based on resource constraints. It suggests that computationally efficient hybrid models are viable for real-time content moderation, whereas deep learning may be reserved for scenarios where resources are abundant and maximum accuracy is prioritized over speed.
4. Identifying Research Gaps: The study highlights specific areas requiring future investigation, including the need for robust multilingual detection frameworks, explainable artificial intelligence (XAI) to improve transparency, standardized evaluation protocols, and the exploration of multimodal detection (integrating text, image, and audio) to address sophisticated misinformation.

In summary, the paper argues that the path to robust fake news detection lies not solely in adopting the most complex deep learning architectures, but in the strategic combination of efficient feature representation (specifically hybrid TF-IDF and N-gram), ensemble learning, and appropriate classifier selection tailored to the specific constraints of the application environment.