VGCas: distinguishing the cascade structure and the global structure in popularity prediction

Online social platforms like Twitter, Weibo, and Facebook have developed rapidly in recent years. These platforms offer people more opportunities to exchange information. Understanding and predicting information cascade on social media platforms is a fundamental problem and one of the primary challenges is to predict the popularity of information. However, most existing methods fail to distinguish the cascade structural feature and global structural feature, resulting in unsatisfactory prediction performance. In this paper, we propose a novel framework named VGCas to distinguish the features of cascade structure and global structure and combine them with temporal features of the cascade to predict popularity. To extract the cascade structural feature and global structural feature simultaneously, we utilize a graph attention based variational autoencoder. Then, we use a gated recurrent unit to extract the temporal feature from the time series. Finally, we feed the combination of the two outputs into a multilayer perceptron to predict popularity. We verify the effectiveness of VGCas by applying it to predict retweet cascades on Twitter and Sina Weibo. Experimental results demonstrate a substantial improvement in predictive accuracy over existing approaches.