Information diffusion prediction via meta-knowledge learners

Information diffusion prediction is a fundamental task for a vast range of applications, including viral marketing identification and precise recommendation. Existing works focus on modeling limited contextual information from independent cascades while overlooking the diverse user behaviors during the information diffusion: First, users typically have diverse social relationships and pay more attention to their social neighbors, which significantly influences the process of information diffusion. Second, complex temporal influence among different cascade sequences leads to unique and dynamic diffusion patterns between users. To tackle these challenges, we propose MetaCas, a novel cascade meta-knowledge learning framework for enhancing information diffusion prediction in an adaptive and dynamic parameter generative manner. Specifically, we design two meta-knowledge-aware topological-temporal modules – Meta-GAT and Meta-LSTM – to extract cascade-specific topological and temporal user interdependencies inherent within the information diffusion process. Model parameters of topological-temporal modules are adaptively generated by the constructed meta-knowledge from three important perspectives: user social structure, user preference, and temporal diffusion influence. Extensive experiments conducted on four real-world social datasets demonstrate that MetaCas outperforms state-of-the-art information diffusion models across several settings (up to 16.6% in terms of Hits@100).