An Early Detection Model For Social Media Rumors Considering Users' Emotional Characteristics

Abstract Rumors in social media platform seriously affect social order and daily life. Aiming at the problem of early rumor detection in social media, this paper proposes a rumor early detection model, REDM, which integrates the historical behavior characteristics and emotional characteristics of UGC publishing subject users in social media. Firstly, the user feature index system of social media was constructed from the perspectives of basic features and user historical behavior features F o , semantic features F s and emotional features F e of UGC corpus. Secondly, according to the UGC corpus released by user history, three behavioral characteristics of user history emotional tendency, emotional fluctuation and credit degree are calculated. Then, BERT model and CNN model were used to extract semantic features and emotion features of UGC corpus. Considering the temporal characteristics of UGC propagation, the RNN model was used for sequence modeling. F s ' , F e ' and F o ' obtained by fine-tuning F o , F s and F e respectively are splicing into a single UGC corpus feature F all , the characteristics of F all_1 , F all_2 , …, F all_N of UGC corpus t 1 , t 2 , …, t N are entered into LSTM at each time to realize early rumor detection task. The experimental results show that the detection accuracy of the model proposed in this paper is better than other baseline models. The accuracy rate increased by 1.7% and the recall increased by 3.78%. At the same time, the rumor event detection in the early stage, less relevant posts context can also have a better performance.