The proliferation of fake news on social media platforms poses a significant threat to social cohesion and trust. Reputable news organizations have migrated to social media, but this has also facilitated the spread of misinformation. Fake news, defined as provably false information intended to deceive, impacts political discourse, public health decisions (e.g., vaccination), and financial markets. Manual fact-checking is insufficient to combat this issue due to its limitations in scalability and automation. Machine learning and deep learning offer potential solutions. Transformer-based pre-trained language models, such as BERT, have achieved state-of-the-art results in various NLP tasks, but their application to fake news detection remains limited. This research aims to address this gap by developing a hybrid model that combines the strengths of BERT for deep text representation and LightGBM for efficient classification, offering a faster and potentially more accurate approach to fake news detection.

Existing fake news detection methods fall into traditional machine learning and deep learning approaches. Traditional methods, including Naive Bayes, linear regression, support vector machines (SVM), decision trees (DT), and XGBoost, often utilize techniques like n-gram analysis and TF-IDF for feature extraction. While some studies report high accuracy on specific datasets (e.g., 92% accuracy with SVM on the ISOT dataset), generalizability to other datasets is often questionable. Deep learning methods, such as CNNs and LSTMs, have also been explored, sometimes in hybrid models. However, these methods may struggle with long-range contextual information and context-specific word representation. Recent work has utilized transformer models like BERT, XLNet, RoBERTa, DistilBERT, and ALBERT, achieving comparable results. While BERT-based models show promise, they can be computationally expensive. The proposed method aims to overcome these limitations by integrating BERT with LightGBM, balancing accuracy and computational efficiency.

The proposed method is a hybrid model combining a fine-tuned BERT model with a LightGBM classifier. The process begins with pre-processing steps to clean the input text: removing non-alphabetic characters, URLs, numbers, stop words, punctuation, and applying case folding. Texts with fewer than ten words are excluded. The pre-processed text is then tokenized and fed into a pre-trained BERT model. The [CLS] token embeddings from the last three hidden layers of the fine-tuned BERT model are concatenated to create a feature vector. This feature vector is then used to train a LightGBM classifier, which predicts whether the news is real or fake. The LightGBM model is chosen for its efficiency and scalability, utilizing techniques like histogram-based splitting, leaf-wise tree growth, gradient-based one-side sampling (GOSS), and exclusive feature bundling (EFB) to optimize training speed and memory usage. The methodology is detailed in Algorithm 1. Three datasets—ISOT, TI-CNN, and FNC—were used for evaluation, each split into training, testing, and validation sets. The performance of the proposed method is compared with several other methods using various word embedding techniques (TF-IDF, GloVe, and BERT embeddings) and classifiers (MNB, LR, SVM, LSTM). Evaluation metrics include accuracy, precision, recall, F1-score, and AUC.

The proposed hybrid BERT-LightGBM model consistently outperforms other methods across all three datasets (ISOT, TI-CNN, FNC) for both headline-only and full-text classification. The comparison includes traditional machine learning models (MNB, LR, SVM) and deep learning models (LSTM) using different word embedding techniques (TF-IDF, GloVe). On the ISOT dataset, the proposed method achieves an accuracy of 99.88% for full text classification and 98.91% for title classification, significantly exceeding the performance of TF-IDF+SVM (99.42% and 95.15%), GloVe+LSTM (98.45% and 97.94%), and BERT+LSTM (99.85% and 98.52%). Similar superior performance is observed on the TI-CNN dataset (96.94% accuracy for full text and 88.65% for title) and FNC dataset (99.06% accuracy for full text and 86.38% for title). The analysis of different embedding techniques (CLS, MeanPooling, MaxPooling) within BERT demonstrates that concatenating the [CLS] embeddings from the last three layers (CLS3) yields the best performance. The ROC curves and AUC values further confirm the superiority of the proposed method across datasets. Comparisons with state-of-the-art methods in Table 4 highlight the significant improvement in accuracy and F1-score achieved by the proposed hybrid model.

The superior performance of the proposed hybrid model can be attributed to the strengths of both BERT and LightGBM. BERT's ability to capture complex linguistic patterns and semantic relationships in text, coupled with LightGBM's efficient handling of high-dimensional feature spaces and optimization capabilities, results in a powerful and efficient fake news detection system. While BERT is computationally intensive, its use as a feature extractor, followed by LightGBM's parallel and memory-efficient training, mitigates this issue. The results demonstrate the effectiveness of this hybrid approach, suggesting a promising solution for addressing the growing problem of fake news. The consistent superior performance across diverse datasets indicates the robustness and generalizability of the proposed method.

This paper presents a novel hybrid model for automated fake news detection, combining BERT and LightGBM. The results demonstrate superior performance compared to state-of-the-art methods across multiple datasets. Future work could explore the integration of other features (e.g., author credibility, source reputation, image analysis) to further enhance the accuracy and robustness of the system. Investigating different BERT architectures and LightGBM hyperparameter tuning strategies could also improve performance. The model's adaptability to evolving fake news techniques should also be assessed.

While the proposed method demonstrates high accuracy, the generalizability to entirely new types of fake news or significantly different languages requires further investigation. The reliance on labeled datasets for training is a limitation, as creating large, high-quality labeled datasets can be challenging. The study focused primarily on textual content and didn't incorporate multimodal features (images, videos), which could provide additional context for more accurate detection. Finally, the computational resources required for BERT fine-tuning might pose a barrier for users with limited computational capabilities.