Predicting information diffusion via deep temporal convolutional networks. Issue 108 (September 2022)

• Record Type: Journal Article
• Authors: Zhao, Qihang; Zhang, Yuzhe; Feng, Xiaodong
• Is Part Of: Information systems. Issue 108(2022)
• Publication Date: 2022-09
• Subjects: Social networks -- Information diffusion -- Dynamic mapping -- Temporal convolutional networks; Database management -- Periodicals; Electronic data processing -- Periodicals; Bases de données -- Gestion -- Périodiques; Informatique -- Périodiques; Database management; Electronic data processing; Periodicals; 005.7
• DOI: 10.1016/j.is.2022.102045
• Languages: English
• ISSNs: 0306-4379

Abstract

Abstract: Information cascade diffusion is ubiquitous in modern social medias and other fields, such as viral marketing, paper citation dynamics, and public opinion communication. However, the existing deep learning-based methods to model and predict the growth of information cascades pay much attention to the nodes in the cascades and ignore the overall propagation structure of the cascades. Simultaneously, they are usually of high complexity and low computational efficiency. In this paper, we propose a novel deep learning framework for information cascade predictor, named CasTCN, which can effectively capture the structure dynamic of the information cascades. Firstly, we design a dynamic mapping mechanism, which can represent the overall structure of information cascades an its dynamic evolution. Secondly, we extract the higher-level representation of cascade network through deep temporal convolutional network. Finally, the prediction module is applied to predict the growth scale of information cascades. Since CasTCN is a graph-level method, it has relatively fewer model parameters, so the model training time is also less than other baseline methods. Our experiments on two real-world datasets show that CasTCN can achieve better performance than other baseline methods on both effectiveness and efficiency. Compared with the state-of-the-art methods, the prediction error is reduced by 8.58%, 7.13% and 6.89% respectively on 3 subdatasets of Weibo, and 15.00%, 11.02% and 9.12% on 3 subdatasets of APS respectively. Also, the training time is much more less than baselines. Highlights: It proposes a novel deep learning framework for popularity prediction: CasTCN. We design a novel mechanism to capture the dynamic structure information of network. CasTCN only needs structure and temporal information of networks. CasTCN can achieve state-of-the-art in two real-world datasets.