Cascade Popularity Prediction: A Multi-View Learning Approach With Socialized Modeling

Predicting the popularity of information in social networks poses a highly challenging problem. The popularity of a message is contingent upon its diffusion process and the relationships it maintains with other cascades and is influenced by user behaviour within the social network. Effectively capturing the dynamic process of message propagation and integrating the structural features of the social network to enhance popularity prediction constitutes a pivotal challenge. To address the challenge, we propose a novel method called “Cascade Social Net” (CSN) that leverages cascade graphs and social graphs to predict cascade popularity accurately. The proposed method consists of three stages. Firstly, we construct a social graph by collecting user information and their connections. Secondly, we integrate information from social graphs, cascade graphs and inter-cascade graphs. Finally, we leverage graph neural networks to predict the popularity of cascades. To overcome the challenge of large-scale social graphs, we introduce a novel neighbour sampling technique that efficiently aggregates information from second-order neighbours. We evaluate our method on real-world datasets and compare it with state-of-the-art methods. Our results demonstrate that CSN outperforms existing methods in predicting cascade popularity.