Cross-modal deep fusion based on small samples for early rumor detection

Rumors circulating on social media can have significant adverse impacts on society, highlighting the urgent need for effective rumor detection. However, most existing methods predominantly focus on rumor identification only after widespread dissemination has occurred, when substantial harm has been inflicted. Early-stage rumor detection is a major challenge due to limited reach and small sample sizes, which restrict the use of large datasets and traditional propagation models. To overcome these challenges, a cross-modal deep fusion method based on small samples for early rumor detection is proposed, which includes a multimodal feature extraction network and a cross-modal deep fusion network. The multimodal feature extraction network captures features from multiple modalities, while the multimodal deep information extraction network derives deep representations from these modalities. The cross-modal deep fusion network integrates textual and visual features for rumor classification. Furthermore, an enhanced meta-learning training approach based on model-agnostic meta-learning is proposed to improve the efficiency of rumor detection by employing distinct learning rates both within and between tasks. Experimental results on two publicly available datasets demonstrate that the proposed cross-modal deep fusion method outperforms baseline methods and exhibits promising performance.