An Information Diffusion Prediction Model Aligning Multiple Propagation Intentions With Dynamic User Cognition

As a fundamental task in understanding the information diffusion process, information diffusion prediction has garnered significant attention in recent years. However, most existing studies tend to focus on the structural characteristics of information diffusion while neglecting an important phenomenon: the propagation of topics often carries multiple intentions, which align with specific user cognition at different evolutionary stages. This diversity in propagation intentions and the dynamic nature of user cognition pose challenges for prediction tasks. To address the above issues, this article introduces Buzz, an information diffusion prediction model, by innovatively approaching the problem from the perspective of aligning multiple propagation intentions with dynamic user cognition. First, to tackle the dynamic hierarchy of propagation intentions, a concise and efficient cascade intention extraction module is designed. This module uses observed diffusion cascades as intention anchors and employs an improved self-attention mechanism to generate representations of the current multiple propagation intentions. Based on this, attention weights are utilized to dynamically stratify the hierarchical structure of propagation intentions. Second, to address the dynamic nature of user cognition, we take social relationships as cognition anchors to initialize the latent diffusion network and dynamically weight the adjacency matrix through temporal slicing. This accurately models the dynamic diffusion process of topics. On this foundation, the cascades are segmented along the diffusion timeline into corresponding time slices, and graph convolution is applied to refine user dynamic cognition. Finally, considering the complexity of aligning propagation intentions with user cognition, we design a multihead attention fusion module. This module aligns propagation intentions based on user cognition, enabling more precise selection of target users. The great performance on four public datasets validates the effectiveness of our proposed approach.