A Predictive Model for Information Diffusion Combining Individual Association and Group Influence

Abstract

Micro-level prediction of information diffusion aims to predict the next user to participate in the diffusion process, and it is an important task in the field of social network analysis. However, the existing research has two main issues. On one hand, they rely solely on social relationships to learn users' social homophily, leading to an insufficient capture of complex diffusion relationships among users. On the other hand, they overlook the impact of group influence on cascade diffusion, which limits predictive performance. To address the above issues, this study proposes a predictive model for Information Diffusion combining Individual Association and Group Influence, denoted by IGIDP. First, a user diffusion association graph is constructed based on cascade sequences, using a Graph Convolutional Network (GCN) to learn users' structural features. A gated fusion mechanism is then employed to enhance feature representation for better learning of the impact of user diffusion relationships on cascades. Next, a hypergraph is built through user-cascade interactions, and a hypergraph attention network is introduced to learn users' global interaction feature representations. Then, a novel Transformer variant is designed to capture both individual user and group effects on cascade diffusion. Finally, a decoder provides the diffusion probability for each user. Experimental results on four public, real-world datasets show that the IGIDP model achieves improvements in Hits@$k$ and Map@$k$ by 0.42%–20.94% and 1.66%–25.20%, respectively.