Intelligent Prediction Model for Key Group Propagation Based on Community Partition and Multiple Influences

Predicting the communication behavior of key groups in guiding topics plays an important role in aspects such as advertising marketing and rumor supervision. Against the backdrop of the rapid development of Large Language Models (LLMs), aiming at the problems of uncertain topic communication scale and diverse influence drivers in social networks, an intelligent prediction model for key group communication based on community division and multiple influences is proposed. First, starting from the breadth of the communication space, to address the uncertainty in the scale of topic dissemination, we propose the community intelligent detection MB-Link algorithm to segment different groups. This algorithm quantifies attributes such as the number and size of communities where different groups are located, so as to characterize the scale of communication. Additionally, leveraging the excellent graph processing capability of Graph Convolutional Networks (GCN), we construct node representations based on the breadth-oriented communication network. Second, in view of the differences between network structure information and text information, an LLM-driven ST2vec representation method is proposed. It extracts the influence of topic dissemination from two levels: user individuals and their friends, and quantifies various influencing factors to reconstruct the topological structure of topic dissemination. Meanwhile, this method combines the sentiment analysis capability of LLMs to determine users’ emotional tendencies, thereby improving the accuracy of user behavior prediction. Finally, in response to the timeliness and dynamics of topic dissemination, time slicing is used to discretize the dissemination data. An intelligent prediction model for key group communication based on community division and ST2vec is proposed, which combines the CNN network to predict the behavior of key groups at the next dissemination moment. Experiments show that this model can not only effectively divide key groups in different communities, but also accurately predict group communication behaviors in topic networks based on complex influencing factors.