CasHyper: Modeling multi-factor interactions and asymmetric influence in cascade prediction via hypercomplex algebra
Year: 2026
Type: article
Abstract: Cascade prediction aims to forecast the future evolution of information diffusion and is an important task in social network analysis. Most existing methods represent users and cascades in real-valued Euclidean spaces, which may offer limited capacity to model multi-factor interactions and asymmetric influence in complex diffusion processes. To address these challenges, we propose CasHyper, a hypercomplex-based cascade prediction framework that integrates hypercomplex algebra with graph neural networks and Transformers. CasHyper organizes user and cascade embeddings into multiple hypercomplex components, with each component encoding distinct latent factors, and uses structured algebraic operations to model cross-component interactions. The non-commutative hypercomplex product further yields order-dependent interaction representations, enabling the model to learn asymmetric influence among users. In the network encoding stage, a static heterogeneous graph channel captures user-user and user-cascade relations, while a Neural ODE-based dynamic channel models the continuous evolution of node states and edge weights. A hypercomplex Transformer with norm-preserving rotary transformations is then introduced to incorporate relative positional information into cascade sequence modeling. Finally, an interaction-aware module and a linear projection module are combined to estimate the probability of each candidate user participating next. Experiments on four real-world datasets show that CasHyper achieves competitive or better performance than the compared methods under the evaluated settings, including relative improvements of up to 43.24% on Android and 18.43% on Christianity in Hits@50. Additional analyses of architectural variants, graph incompleteness, computational efficiency, and different Cayley–Dickson algebras further demonstrate the empirical effectiveness of the proposed framework and reveal dataset-dependent behaviors of different hypercomplex structures. We release the code at https://github.com/MrYansong/CasHyper .
Authors: Yansong Wang, Qisen Chai, Junjie Huang, Tao Jia
Institutions: Southwest University, Chongqing Normal University
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,670
OpenAlex ID: W7222792193