Abstract
Information cascade prediction has demonstrated wide application value in public opinion guidance and social diffusion modeling. However, existing approaches suffer from limited capabilities in modeling long-range temporal dependencies, lack unified temporal representations that account for varying user activity patterns, and often overlook the imitation-driven motivations underlying user diffusion behavior. To address these limitations, we propose imitation-based multiscale framework for cascade prediction (IMF-Cas), a cascade prediction framework integrating behavioral imitation with dynamic graph representation. The framework introduces a social time transformation to normalize natural time into an activity-aware timeline, mitigating bias from uneven user activity distributions. A sparse graph convolutional network (GCN) encodes local structural features at each time slice, while a transformer with sparse attention captures temporal dependencies across slices for efficient long-range sequence modeling. We design a multiscale imitation mechanism that characterizes users' reposting motivations through micro-level features (trust relationships and interaction intensity) and macro-level features (local imitation ratio and clustering coefficient). A dynamic information hotness function combining semantic properties with temporal decay is integrated to model the evolving attractiveness of content. Experiments on the Twitter and Weibo datasets show that IMF-Cas reduces mean squared logarithmic error (MSLE) by 35.3% and 22.9% relative to the strongest competing baseline on each dataset, with consistent gains on micro-level tasks, validating its effectiveness in capturing cascade dynamics. These results also point to practical cascade control strategies, including early warning systems and targeted interventions for rumor suppression.