Topic-Aware Multi-heterogeneous Graph Neural Network for Predicting the Next Participant in Information Cascades

Yichao Zhang, Zejian Wang, Guanghui Wen, Jihong Guan, Shuigeng Zhou, Wengen Li, Zhiwei Cao

Abstract: Predicting the next participant in an information cascade on social networks has always been an open problem. Recently, graph neural networks (GNN) have been demonstrated to be one of the most effective schemes to handle this task. The GNN improves users’ representations by aggregating the representations of the users followed directly or indirectly by them in the social network. In contrast, these links are only one type of pathway for users to receive information. There are yet information propagation channels between disconnected users. Previous studies on the information propagation channels are limited, let alone a deep learning framework utilizing them for participant prediction. Inspired by these facts, we first mine the hidden information propagation channels from users’ tweets in cascades and construct a multi-heterogeneous graph by combining these channels with the following links. Subsequently, we propose a topic-aware multi-heterogeneous graph neural network (TMGNN) model to exploit the extended channels for learning user representations. We learn the attention among users’ extended neighbors on a particular topic and integrate topic-aware attention to tackle the task in cascades involving unknown topics. Extensive experimental results on two publicly available datasets show that the TMGNN achieves a relative improvement of over 15% against the best performance of the tested state-of-the-art models in terms of MAP@10, demonstrating its generalizability and robustness. Code for this study is publicly available at https://github.com/william-wang-stu/TMGNN.