Broad information diffusion modelling for sharing link click prediction using knowledge graphs

Abstract

In the new media era, users actively share and diffuse information across social networks, creating complex patterns of broad information diffusion (BID) that differ significantly from traditional recommendation scenarios. Existing models are primarily designed for deep information diffusion (DID) with sequential cascades and struggle to address BID challenges, including the sparse graph structure, weak temporal correlation, and ambiguity in user preferences. To bridge this gap, we propose K-BID, a knowledge-driven framework tailored for BID scenarios. K-BID integrates semantic and social graph information through a two-phase ‘Match & Rank’ approach. The matching phase retrieves candidate voters using social relationships and personalized preferences, whereas the ranking phase refines predictions by modelling temporal dynamics. Experiments on real-world datasets demonstrate the superiority of K-BID over state-of-the-art methods, achieving significant improvements of 14.02%, 16.80%, and 16.99% in Precision, MRR, and AUC respectively, for the ‘Soc.’ objective with $K=5$. Our work advances the understanding of BID scenarios and offers a practical solution for optimizing information dissemination in social platforms.