Adaptive Weighted Ensemble Deep Learning for Robust Rumor Detection on Social Media

Rumor detection in social media is a critical concern and is exacerbated by the complexity and diversity of rumor texts. Within social media rumor posts, auxiliary elements such as real-time emerging comments regarding the event can reveal public uncertainty and skepticism about the content in question. Motivated by this intuition, we propose the Adaptive Weighted Ensemble Deep Learning model (AWEDL), an innovative framework that integrates rumor and stance models. AWEDL adeptly captures user attitudes embedded in comment data, without requiring explicit stance labels. Through a comprehensive process of weighted fusion and dynamic adjustment of vector representations pertaining to rumors and stances, AWEDL strikes an optimal balance between these features. This approach culminates in an increase in the precision of rumor detection. Extensive validation on three well-recognized datasets, including Twitter15, Twitter16, and Weibo, firmly establishes AWEDL's superiority over existing benchmarks, with notably remarkable performance on the Weibo dataset, achieving an average F1 score of 95.7%.