Adaptive sparse graph learning for interpretable cross-modal misinformation detection

Multimodal misinformation detection requires understanding complex relationships between text, images, and metadata. While existing fusion methods often introduce noise through dense connections or fail to capture discriminative cross-modal patterns, we propose AdaSGL (Adaptive Sparse Graph Learning), a novel framework that dynamically constructs sparse heterogeneous graphs to model task-relevant multimodal interactions. Our key insight is that misinformation often manifests as specific inconsistencies between modalities—yet most connections in dense graphs are irrelevant noise. AdaSGL addresses this through three innovations: (1) a learnable sparsity mechanism that adaptively prunes edges based on their contribution to detection accuracy, (2) modality-aware graph propagation that preserves distinctive patterns while enabling cross-modal reasoning, and (3) a hybrid optimization objective that jointly learns graph topology and classification boundaries. Experiments on three benchmarks demonstrate that AdaSGL achieves state-of-the-art performance (89.7% on FakeNewsNet, 84.6% on TwitterMisinfo, 82.4% on MM-COVID) while reducing computational costs by 35-57% through selective edge retention. The learned sparse graphs provide inherent interpretability by highlighting critical cross-modal relationships, with our Cross-Modal Consistency metric showing 9.9-16.5% improvement in detecting modality inconsistencies.