Modeling topic evolution in public opinion events: an unsupervised spatio-temporal graph attention approach

With the widespread use of online social media, Public Opinion Events (POEs) quickly propagate on the Internet, generating a vast amount of textual data centered around various discussed topics. The development of POEs is closely linked to the evolution of these topics. However, in developing of POEs, the key challenges lie in estimating the duration of different topics, dealing with their dynamic natures, and quantifying topic evolution to predict the number of topics in the future. In this paper, we propose an Unsupervised Spatio-Temporal Graph Attention approach (USTGAT-TT) to tackle these challenges. First, we introduce a topic evolution periods generation method without human intervention. Initially, POEs data undergoes pre-processing to establish initial periods and extract keywords. According to the persistence and hotness of keywords, new periods are reconstructed and keywords are clustered by their similarity to form topics. Then we analyze three pieces of knowledge to further learn the evolution of topics, macro properties, micro properties and dynamic topic network graphs via topics co-occurrence relationship. Finally, we design a Spatio-Temporal Graph Attention topic trend prediction model (STGAT-TT) by taking the mutual effect of topics and temporal dependencies into account. At the same time, attention mechanism and average method are employed to obtain the contribution of topics and Long Short-Term Memory (LSTM) is used to predict the number of topics in the next period to study the state of POEs. Experiments on five POEs show that the effectiveness of the proposed approach. It can estimate the duration of topics to form periods and quantify their features to learn evolution and predict the number of topics in the next period.