Social opinions prediction utilizes fusing dynamics equation with LLM-based agents

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Network Science

Background:

• Social media platforms are crucial for social movements and public opinion.
• Traditional simulation models often fail to capture real-world social complexities.
• Accurate opinion dynamics simulation is vital for understanding societal trends and informing policy.

Purpose of the Study:

• To propose and evaluate the novel Fusing Dynamics Equation-Large Language Model (FDE-LLM) algorithm.
• To enhance the accuracy of simulating user opinion dynamics on social media.
• To demonstrate the superiority of the FDE-LLM over traditional Agent-Based Modeling (ABM) and existing LLM approaches.

Main Methods:

• The FDE-LLM algorithm models users as either opinion leaders or followers.
• Opinion leaders utilize Large Language Models (LLMs) for role-playing and Cellular Automata (CA) for opinion constraint.
• Opinion followers are simulated using a combined CA and Susceptible-Infectious-Recovered (SIR) model.

Main Results:

• Experiments on four real-world Weibo datasets show FDE-LLM significantly outperforms traditional ABM and LLM algorithms.
• The algorithm accurately simulates the temporal decay and recovery of user opinions.
• FDE-LLM demonstrates improved prediction accuracy for social media opinion dynamics.

Conclusions:

• The FDE-LLM algorithm represents a significant advancement in simulating social media opinion dynamics.
• LLMs, when integrated with dynamic systems like CA and SIR, can revolutionize social science research.
• This approach offers a more nuanced and accurate understanding of how opinions spread and evolve online.