Hybrid Feature Combinations with CNN for Bangla Fake News Classification

Technical Summary: Hybrid Feature Combinations with CNN for Bangla Fake News Classification

Problem Statement
The proliferation of false news on social media and online portals in Bangladesh poses significant risks to the credibility of authentic media and the daily lives of citizens. Detecting fake news in the Bangla language presents unique challenges due to its complex writing structures and diverse vocabulary, which differ from English. While numerous studies exist for English fake news detection, research specific to Bangla remains limited. Existing approaches often rely on single or a few feature extraction techniques (e.g., TF-IDF, Word2Vec, or fastText) without systematically comparing their effectiveness or exploring hybrid combinations. Furthermore, high-dimensional feature sets can introduce noise and reduce model efficiency, necessitating robust feature selection strategies.

Methodology
This research proposes a comprehensive framework for Bangla fake news classification using the BanFakeNews-2.0 dataset, which comprises 48,678 real and 12,903 manually annotated fake news articles. The methodology follows a structured pipeline:

1. Preprocessing: Raw Bangla text is cleaned by removing punctuation, stop words, special characters, and HTML tags, while retaining digits.
2. Feature Extraction: Six distinct feature extraction methods are employed to convert text into numerical representations:
   • TF-IDF: Utilizing unigrams and bigrams (top 5,000 tokens).
   • Word2Vec: Using the CBOW architecture with a 100-dimensional vector space.
   • FastText: Leveraging subword information for better handling of unseen words.
   • N-Gram Features: Extracting sequences up to size 3.
   • Character-Level TF-IDF: Focusing on character n-grams (length 2–4) to capture subword patterns.
   • Statistical Features: Manually constructed descriptors including character/word/sentence counts, average word length, and punctuation ratios.
3. Feature Selection: To identify the most informative features and reduce complexity, five selection approaches are applied:
   • Statistical methods (ANOVA F-test, Chi-Square, Mutual Information).
   • Ensemble-based importance (Random Forest).
   • Recursive Feature Elimination (RFE).
   • Correlation-based selection (Pearson coefficient).
   • Forward Selection and Comprehensive Combination Testing (brute-force evaluation of all subsets).
4. Classification Model: A Convolutional Neural Network (CNN) with a multi-branch architecture is proposed. This design processes different input feature types through separate convolutional pipelines, compresses them via global max pooling, and concatenates the resulting vectors for binary classification (Authentic vs. Fake). The model is trained using Binary Crossentropy loss, the Adam optimizer, and early stopping mechanisms.

Key Contributions

• Systematic Feature Evaluation: The study provides a comparative analysis of six diverse feature extraction techniques specifically for Bangla text, moving beyond the reliance on single-method approaches common in prior literature.
• Hybrid Feature Combinations: The research demonstrates the efficacy of combining semantic (Word2Vec, FastText), statistical (TF-IDF, N-Gram), and structural (Character-level, Statistical) features.
• Rigorous Selection Strategy: By employing multiple feature selection algorithms and a brute-force combination test, the study identifies optimal feature subsets that maximize classification performance while minimizing redundancy.
• Open Source Implementation: The code and experimental setup are made publicly available to facilitate further research in this domain.

Experimental Results
The proposed CNN model was evaluated on the BanFakeNews-2.0 dataset using Accuracy, Precision, Recall, and F1-score.

• Individual vs. Hybrid Performance: Individual features like FastText and Word2Vec achieved comparable performance (F1-score ~0.80). However, combining features yielded significant improvements.
• Optimal Combination: The highest performance was achieved by integrating TF-IDF, Word2Vec, FastText, Character-level TF-IDF, and Statistical features. This hybrid approach resulted in:
  • Accuracy: 0.91
  • Precision: 0.90
  • Recall: 0.79
  • F1-score: 0.83
• Fake Class Specifics: For the specific detection of fake news, the hybrid model improved the F1-score to 0.73 and recall to 0.61, a marked increase over individual features (e.g., Word2Vec alone had a recall of 0.55 and F1-score of 0.67).
• Training Stability: Training curves indicated stable learning behavior with no signs of overfitting across 50 epochs.

Significance and Claims
The paper claims that the primary significance of this work lies in demonstrating that combining multiple feature types significantly enhances the recall and F1-scores of Bangla fake news detection compared to using individual features alone. The authors emphasize that integrating diverse semantic and statistical features allows the model to capture complementary information, leading to more robust classification.

The study concludes that while individual deep learning embeddings perform reasonably well, they often miss a significant portion of counterfeit instances (low recall). The proposed hybrid approach addresses this limitation. The authors modestly note that their study is limited to a CNN classifier and suggest that future work could explore other machine learning and deep learning architectures, as well as additional feature extraction and selection techniques.