CasSampling: Exploring Efficient Cascade Graph Learning for Popularity Prediction

Abstract

Predicting the growth size of an information cascade is one of the primary challenges in understanding the diffusion of information. Recent efforts focus on utilizing graph neural networks to capture graph structure. However, there is considerable variance in the information cascade size (from few to million). From the perspective of efficiency and performance, the method of modeling each node is inappropriate for graph neural networks. In this paper, we propose a novel deep learning framework for popularity prediction called CasSampling. Firstly, we exploit a heuristic algorithm to sample the critical part of cascade graph. For the loss of structure information due to sampling, we keep outdegree of sampled node in the global graph as part of the node feature into the graph attention networks. For the loss of temporal information due to sampling, we utilize the time series to learn the global propagation time flow. Then, we design an attention aggregator for node-level representation to better integrate local-level propagation into the global-level time flow. Experiments conducted on two benchmark datasets demonstrate that our method significantly outperforms the state-of-the-art methods for popularity prediction. Additionally, the computation cost is much less than the baselines. Code and (public) datasets are available at https://github.com/Gration-Cheng/CasSampling .