A Dynamic Short Cascade Diffusion Prediction Network Based on Meta-Learning-Transformer

The rise of social networks has greatly contributed to creating information cascades. Overtime, new nodes are added the cascade network, which means network is dynamically variable. At same time, there often only a few in before join. Therefore, it becomes key task predict diffusion after dynamic based on small number observed previous period. However, existing methods limited for short cascades and cannot combine temporal with structural well, so model, MetaCaFormer, meta-learning Transformer structure, proposed this paper prediction. Considering processing capability traditional graph neural information, we propose CaFormer model inherits powerful while considering neighboring nodes, edges spatial importance effectively combining information. improve prediction ability cascades, also fuse that can be quickly adapted data. In paper, MetaCaFormer applied two publicly available datasets different scenarios experiments demonstrate its effectiveness generalization ability. experimental results show outperforms currently baseline methods.