Predicting Participation Shift of Users at the Next Stage in Social Networks

Abstract

In online social networks, numerous studies have demonstrated the challenge of predicting who will eventually engage in an information cascade with its initial part. Take a step back. Can we predict who will engage in the cascade at the next stage if the lifetime of cascades is divided into a certain number of stages? Although numerous attempts have been made to solve this problem, how to extract useful information from the historical cascades spreading within a sub-network and the connections among users remains an open question. This paper proposes a simple but efficient unsupervised agent-based model, the triple ranking model, which integrates exposure time ranking, social gravity ranking, and cascade similarity ranking. The rankings, a key component of our model, have been successful in characterizing the social impact of shifted users, temporal information, and sequential cascade information, demonstrating the generalizability of our approach. To test the contributions of the features in supervised frameworks, we fuse them with two graph neural networks, the graph convolutional network (GCN) and graph attention network (GAT). Our experimental results on three Twitter networks unequivocally show that the proposed algorithm outperforms the tested state-of-art algorithms across a series of performance metrics. Notably, its time complexity is also lower than theirs, further underscoring its superiority. The observations demonstrate that the rankings effectively abstract the features hidden in the information cascades and in the topology of social networks, paving the way for further studies on posting engagement.