Multimodal Meta-Learning for Early Rumor Detection with Adaptive Decision Timing

Abstract

Early social media rumor detection is essential because it prevents false information from spreading quickly and causing serious harm. Current methods frequently ignore the significance of prompt decision-making in favor of increasing classification accuracy. In this paper, we present a novel framework for early rumor detection that is based on multimodal meta-learning with adaptive decision timing. The method analyzes social media posts as an ongoing stream and learns when enough data has been observed to dynamically identify the best time to make a forecast. By integrating textual, structural, and temporal features, the proposed approach captures complex rumor propagation patterns while jointly optimizing accuracy and detection timeliness. Experimental results on benchmark datasets, including Twitter15, Twitter16, and Weibo, show that the proposed method significantly reduces detection latency, requiring substantially fewer messages for prediction, while maintaining competitive or superior performance compared to state-of-the-art baselines.