TCDI-Net: Early fake news detection network via topical context enhancement and cross-modal deep interaction

To address the challenges of sparse information and high deceptiveness in early detection of fake news on social media, where visual and textual elements appear consistent but narrative frames are misaligned, we propose the Early Fake News Detection Network via Topical Context Enhancement and Cross-Modal Deep Interaction (TCDI-Net). Guided by communication framing theory, the model aims to achieve more accurate discrimination by coordinating explicit content representations with implicit narrative intentions. First, we construct a semantically enhanced topic modeling module that treats global topics as operational representations of narrative frames, providing macro-level topical supplementation for information-sparse early stages. Second, we design a cross-modal deep interaction network. It employs a bidirectional enhanced attention mechanism and a hierarchical fusion structure to capture deep logical connections between text and images at fine-grained levels. Additionally, a gating mechanism and supervised contrastive learning constraints are introduced to dynamically regulate modal contributions and optimize the feature space distribution, thereby improving the model’s ability to identify highly deceptive samples. Experimental results on the public PHEME and Weibo datasets demonstrate that TCDI-Net achieves accuracies of 90.21% and 91.06%, respectively, outperforming existing baselines in early detection scenarios and validating the effectiveness of topical context enhancement and deep interaction mechanisms.