Constructing a multi-layer heterogeneous networks model to explore the public opinion evolution pattern of key users in public health emergencies

This article aims to discover key users in public health emergencies and explore their public opinion evolution patterns, thus providing theoretical support for the establishment of a clear cyberspace. In this article, a multi-layer heterogeneous network model based on multiple node attributes (user nodes, microblogs nodes, emotional nodes and topic nodes) is constructed at each stage of the public opinion cycle. Specifically, a novel semi-supervised self-training method based on the bidirectional encoder representations from transformers (SSST-BERT) method is proposed to automatically label the fine-grained emotional nodes. The latent Dirichlet allocation (LDA) model is used to construct the topic nodes. Moreover, the degree centrality, betweenness centrality and closeness centrality of the constructed heterogeneous network are adopted to dynamically identify key users. Finally, the emotional states and topic tendencies of key users are explored to obtain the public opinion evolution pattern of emergencies. The experimental results show that the SSST-BERT automatically labels emotion categories with an F1-score of 80.48%. The key users identified by the constructed heterogeneous network are more representative of the opinion status of ordinary users. Analysis of the public opinion status of key users reveals that netizens show more negative emotions such as anger and fear in public health emergencies, and the shift of focus drives the evolution of discussion topics.