A Hybrid CNN-BiLSTM Framework with Attention-Based Explainability for Interpretable Fake News Detection

Abstract

The rapid spread of fake news on social media has far-reaching implications, affecting various aspects of national life, including political stability, governance, economic systems, public health, and education. These platforms not only facilitate information exchange but also shape global public opinion. Effective fake news detection on social media is therefore essential to protect democratic processes, maintain public trust in institutions, prevent election manipulation, mitigate social polarization and violence, reduce the spread of health misinformation during crises (such as pandemics), curb financial fraud, and preserve the integrity of public discourse. Despite the ongoing efforts in combating fake news, most existing solutions remain limited in scope and interpretability. This study evaluates deep learning (DL) models, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and a hybrid DL model (CNN-BiLSTM with attention mechanism) to improve both robustness and prediction interpretability. The proposed model (CNN-BiLSTM+Attn) was evaluated on a Nigerian social media news dataset (FN_data, 126,974 records after deduplication and preprocessing) across five independent training runs with different random seeds (42 - 46), yielding a mean accuracy of 82.86% ± 0.22% and mean F1-score of 77.12% ± 0.52% as the best performing model, with statistically significant improvement over the other models (p < 0.05). The proposed model was further validated on a standard benchmark dataset (The ISOT fake news dataset), achieving mean accuracy of 99.33% ± 0.10% and mean F1-score of 99.39% ± 0.09% over ten independent runs (seeds 42 - 51). This improvement on ISOT confirms the robustness of the proposed model. Attention visualizations provide token-level explainability, highlighting the model’s focus on deceptive cues. This work provides a transparent