MCDF: Multimodal information fusion and causal analysis for election misinformation detection

The rapid spread of election-related misinformation on social media poses a serious threat to public trust, democratic decision-making, and social stability. This form of misinformation is particularly persuasive and difficult to detect as it uses different types of content (modalities), including text, images, captions, and social interactions. These challenges undermine efforts to ensure trustworthy elections and enable timely intervention by policymakers and fact-checkers. However, existing detection approaches struggle with feature misalignment, cross-modal inconsistencies, and noisy social data, thereby limiting their ability to accurately classify misinformation and explain its propagation. To address these challenges, we propose MCDF, a Multimodal Causal Detection Framework, integrating fusion-driven misinformation detection with causal analysis. Our framework consists of three key components: (1) a multimodal rumor detection module, which employs Graph Convolutional Networks (GCNs) for social interaction modeling, Vision Transformers (ViTs) for visual feature extraction, and RoBERTa for text-caption encoding, dynamically aligned via Tensor Fusion Networks (TFNs); (2) a Noise-Gating Mechanism, which refines feature alignment by filtering misleading or redundant inputs, ensuring robust misinformation classification; and (3) DEMATEL, a causal inference module that quantifies misinformation drivers, bridging misinformation classification with explainability. We evaluate our model on Twitter (X), FakeNewsNet (GossipCO and PolitiFact), and a curated Ghana-specific election dataset, demonstrating state-of-the-art performance in both classification and causal inference. MCDF offers a practical and interpretable framework for combating misinformation in real-world political communication, providing actionable insights for electoral stakeholders, fact-checkers, and social media analysts.