CasTformer: a novel cascade transformer towards predicting information diffusion (Q6052647)

This paper presents a novel cascade transformer towards predicting information diffusion called CasTformer. A global spatio-temporal positional encoding adapted to cascade information in CasTformer is proposed to effectively mark the spatial and temporal position of a node. Based on the cascade graph Laplacian matrix eigenvectors, the global spatial position information is encoded. The paper designs three bias matrices in CasTformer to reflect the relative relationships between nodes in the cascade. A relative interpersonal relationship bias is introduced by creating a triangular relationship among the source node, target node, and their least common ancestor. CasTformer's performance is enhanced by leveraging the self-knowledge distillation mechanism. Numerical examples are provided to illustrate the performance.