Research On Popularity Prediction Of Online Contents In Social Networks

The emergence of online social network has completely changed the way of information diffusion,profoundly affected the information dissemination mechanism and greatly promoted the generation and propagation of information on the Internet.Popularity is a manifestation of the concentration of users' attention,which can effectively reflect the dissemination of information.The ability to predict the popularity of online content accurately has important theoretical significance and broad application prospects,and it has been widely concerned by academia and industry.However,the challenge of this problem comes from the inequality of online content and numerous complex factors.Existing works fall into three main paradigms: feature-driven approaches,generative models,and methods based on deep learning,each with known strengths and clear limitations.In this paper,we carefully analyze the underlying mechanism and various factors governing the popularity dynamics of information cascade.Then,deep learning technology is utilized to model the process of information diffusion and characterize the growth of popularity accurately.The main contributions are as follows:(1)In this paper,we propose a novel deep community-guided structural-temporal convolutional network for popularity prediction.Based on the theory of community detection,we put forward a hypothesis that information diffusion is treated as the inflow and outflow among different communities.We employ convolutional neural networks to model local and global structural dependencies among users from the mesoscopic perspective of communities,and then predict the popularity of online content.The proposed model also provides us great insights in understanding the fundamental mechanism of information diffusion and sheds light on the collective attention on online social networks.(2)We propose a graph attention based spatial-temporal network for popularity prediction.The proposed model can directly learn the representation of information cascade,leveraging an end-to-end deep learning framework to predict the popularity of online content.Graph attention networks and temporal convolutional networks are exploited to extract the structural and temporal features of information cascade,respectively.Experimental results show that its prediction performance is significantly better than comparison models,and it can be widely applied to different data scenarios.(3)This paper presents a context-aware cascade representation learning framework for popularity prediction.Self-excitation mechanism is employed to aggregate the neighbor representation for each user.Then,we exploit two non-parametric time effects to characterize the dynamic process of information diffusion from the micro perspective,and learn the representation of information cascade to predict popularity.The proposed method effectively integrates different information dissemination mechanism and regards the final popularity as supervision,so that it has high interpretability and predictability.