Multimodal Meta-Learning for Early Rumor Detection Based on Few-Shot Learning

Rumors on social media platforms have a significant negative impact on society, making rumor detection increasingly critical. However, most existing methods focus on identifying rumors only after they have already spread widely and negative impacts have occurred. Therefore, identifying rumors in the early stages is necessary. Early rumor detection is typically characterized by limited spread and small sample sizes, making it impractical to rely on large datasets or rumor propagation structures. To overcome these characteristics, a multimodal meta learning method based on few-shot learning (FSL) for early rumor detection is proposed. A multimodal feature extraction layer is designed to extract data features of various modalities, while a multimodal hidden information extraction layer is constructed to uncover deep information from these features. Furthermore, a multimodal fusion output layer is developed to combine and process the multimodal information, leading to rumor classification. A meta-learning algorithm is introduced to address the challenge of small sample sizes, utilizing fast and multistep update methods to enhance the adaptability and stability of model. Comparative experiments conducted on two publicly available datasets confirm that our proposed method demonstrates strong performance in early rumor detection.