Description
In social media, users' opinions are directly influenced by their personal experiences and/or implicitly influenced by the opinions of others in the social network. However, existing public opinion prediction models fail to fully consider the social relationship between users and their temporal dynamics. Opinion spreading is complex and dynamic, and prediction models need to able to accommodate the change patterns in both the temporal and spatial dimensions at the same time. In response, this research aims to address two key scientific questions: (1) how to effectively capture user interactions in complex social networks, and (2) how to incorporate the time factorinto opinion prediction models. In this work, we propose a model based on time-series graph neural networks and design an improved loss function, which can effectively optimize the time-dynamic modeling capability and enhance the prediction effect on long time-span data. The experimental results show that our model performs better than the control model on Weibo and TikTok data, and reveals the differences in opinion propagation between the two platforms. The improved loss function is particularly for graph data in modelling social phenomenon with temporal-spatial characteristics, providing a new approach to dynamic event analysis and showing the potential for a wide range of applications in areas such as opinion monitoring and dynamic
recommendation.
| Period | 1 Jan 2024 → 24 Dec 2024 |
|---|---|
| Degree of Recognition | International |