IGPC: Multi-domain multimodal fake news detection via Info-Gini Weight Adaptation and Prototypical Multi-level Contrastive Learning

Abstract

Multi-domain multimodal fake news detection requires models to integrate textual and visual evidence while generalizing across diverse domains. Existing methods still struggle with multi-level data imbalance, weak class geometry in data-scarce domains, and transfer to held-out target domains. We propose I nfo- G ini Weight Adaptation and P rototypical Multi-level C ontrastive Learning (IGPC), a method that decomposes imbalance into inter-domain, intra-domain, and global levels, and converts dataset statistics into stable training weights. It then learns a domain–label prototype space and applies multi-level contrastive learning to enforce compact intra-class clusters and clear inter-class margins. Finally, a prototype-based transfer enhancement, including a global class-prototype bridge and prediction fusion, improves robustness on held-out target domains without introducing extra loss terms. Experiments on Weibo (7723 experimental samples), Weibo-21,(6156 experimental samples), and FineFake (16,796 experimental samples) show that IGPC outperforms the strongest directly reproduced baseline by 1.06%, 1.48%, and 1.01% in overall F1 score, respectively. It also shows consistent gains in held-out target-domain evaluation on Weibo and Weibo-21 and visible class separation in the learned representations These results indicate that explicitly modeling multi-level imbalance and prototype geometry is effective for robust multi-domain multimodal fake news detection under domain shift.