Modelling the Latent Semantics of Diffusion Sources in Information Cascade Prediction

Figure 3

The architecture of the co-attention fusion gate. The fusion gate is composed of embedding layer, co-attention layer, and fusion layer. (1) The embedding layer transforms diffusion sources and users into two different embedding spaces. (2) The co-attention layer extracts the interaction between users and diffusion sources with bilinear attention mechanism. (3) The fusion layer aggregates the semantics of users and diffusion sources together in support of the further cascade prediction.