Abstract
Detecting fake news on social media is a critical challenge due to its rapid dissemination and potential societal impact. This paper addresses the problem in a realistic scenario where the original tweet and the sequence of users who retweeted it, excluding any comment section, are available. We propose a Graph-based Attention for Coherent Explanation (GRACE) to perform binary classification by determining if the original tweet is false and provide interpretable explanations by highlighting suspicious users and key evidential words. GRACE integrates user behaviour, tweet content, and retweet propagation dynamics through Graph Convolutional Networks (GCNs) and a dual co-attention mechanism. Extensive experiments conducted on Twitter15 and Twitter16 datasets demonstrate that GRACE out-performs baseline methods, achieving an accuracy improvement of 2.12% on Twitter15 and 1.83% on Twitter16 compared to GCAN. Additionally, GRACE provides meaningful and coherent explanations, making it an effective and interpretable solution for fake news detection on social platforms.