Content Matters: A GNN-Based Model Combined with Text Semantics for Social Network Cascade Prediction

Abstract

Effectively modeling and predicting the size of information cascades is essential for downstream tasks such as rumor detection and epidemic prevention. Traditional methods normally rely on tedious hand-crafted feature engineering efforts, which is inefficient in complex diffusion processes such as social network (SN) cascades. In recent years, graph neural network methods have been successfully used in cascade prediction tasks. Most of these methods make use of the structural features of SNs, while the effect of textual user-generated content (UGC) is far from clear or fully utilized. In this paper, we focus on the questions of how the textual UGC affect user activation state and trigger their retweet behaviors. We propose a novel GNN-based model named TSGNN, which jointly model the textual content and structure features. It uses recurrent neural networks with attentions to learn content feature representations that potentially affect information propagation. We find that tweets of fewer high coherent topics are more likely to trigger user retweet behaviors, and we also design a gate mechanism to model the activation state of users under the combined influence of content, structure, and other self-activation. Experimental results demonstrate that TSGNN significantly outperforms all the state-of-the-art methods in multiple metrics.