去中心化的微博传播动力学建模

Decentralized cascade dynamics modeling

社交网络的功能正逐步由网络社交转变为社交媒体,极大地方便了消息的传播,也使得消息的流行度预测问题变得更具有挑战性.传统的流行度预测方法包含基于特征的有监督学习方法和基于随机过程的传播动力学建模方法.其中,基于随机过程的传播动力学建模方法由于具有更好的个体预测能力,吸引了研究人员的广泛关注.但是,现有的传播动力学建模方法在建模时,都忽略了社交平台中消息传播所呈现出的去中心化特点.本文以微博平台中消息的传播数据为基础,分析了微博消息的去中心化传播现象,并提出了一种叠加自增强泊松过程(reinforced Poisson process, RPP)模型的方法来刻画消息的传播动力学变化.每一条信息的传播过程都被拆分为几个传播子过程的叠加,而每个传播子过程可以用RPP模型很好的建模.在真实数据集上的结果表明,本文所提出的方法在刻画消息的传播过程以及预测消息的流行度变化等方面,都要优于现有的方法.

Social network platforms have gradually transformed into social media in recent years, which greatly eases information diffusion. Meanwhile, it also makes the task of popularity prediction for online information more challenging. Traditional popularity prediction methods include supervised methods based on feature engineering and popularity dynamics modeling methods based on stochastic processes. The latter group of methods has drawn a wide attention as it works better for predicting individual items＇ popularity. However, the existing methods have ignored the decentralization characteristic of diffusion cascades. In this paper, we investigate the decentralized structure of diffusion cascades, based on the Weibo dataset, and propose to model cascade dynamics by mixing a bunch of reinforced Poisson processes（RPP）. The overall diffusion cascade is a mixture of several sub-cascades,each of which can be modeled by RPP. Experimental results on real world datasets demonstrate our proposed model＇s superiority in both characterizing popularity dynamics and predicting future popularity.