SRACas: A Social Role-Aware Graph Neural Network-Based Model for Popularity Prediction of Information Cascades

Abstract

Popularity prediction of information cascades is a fundamental and challenging task in social network data analysis. Social roles impact users’ behaviors and change the structure and popularity of information cascades. Existing deep learning-based methods utilize several independent sub-cascade graphs or paths to learn cascade representations, which lose vital information about social roles and dynamics between sub-cascades at different moments. We propose a social role-aware cascade (SRACas) model that exploits the social influences of nodes on previous and subsequent sub-cascade graphs within an observation window to facilitate the social role learning of nodes. A temporal-aware differential loss is also proposed to discriminate the structures of neighboring sub-cascades and captures the dynamics of sub-cascades. Under the techniques of local graph attention, social role-aware attention, and temporal-aware loss, SRACas learns a better latent representation of cascades at both the node level, sub-cascade level, and cascade level. Moreover, there lacks a platform with standard prepossessing procedures that allow convenient configuration and fair competition between information cascade prediction models. An open platform OpenCas is built with uniform preprocesses to verify the faithful performance of the compared methods. Extensive experiments show that SRACas achieved significant improvements over existing methods on classic real-world datasets.

Z. Huang and Y. He—Equal Contribution.