Cascade Prediction with Recurrent Neural Networks and Diffusion Depth Distributions

A message suddenly in the spotlight can be an innocuous superstar scandal, but it can also be an extreme opinion that destabilizes society or a rumor that distorts the truth. Accurately predicting the future cascade size of information cascade on social networks such as Twitter and Weibo can help focus more attention on suspicious large-scale cascades in advance, assisting managers in making decisions for human intervention. However, early cascade prediction methods either rely on a priori stochastic process assumptions or handcrafted features, leading to limited prediction performance. Moreover, most existing deep learning-based cascade prediction methods use graph neural networks or graph representation learning techniques to model cascaded graphs of information propagation, which faces the dilemma of low graph embedding efficiency. To address these issues, considering the strong positive near-linear relationship between the final cascade size and the diffusion depth, we propose a simple yet effective cascade prediction method called Cas3D that captures the evolution of the diffusion depth distribution via recurrent neural networks and models the complex relationship with future cascade sizes. Extensive experiments on two real-world datasets demonstrate that Cas3D significantly outperforms state-of-the-art approaches with a satisfactory efficiency advantage.