PopSim: Social Network Simulation for Social Media Popularity Prediction

TL;DR

PopSim introduces a simulation-based paradigm for Social Media Popularity Prediction by deploying an LLM-based multi-agent social network sandbox to model dynamic UGC propagation. A dual-channel social-mean-field mechanism captures both textual context and agent opinions, while a multi-source information aggregation module unifies heterogeneous metadata for a unified prediction. Empirical results on two real-world datasets show consistent improvements over state-of-the-art inductive methods, with notable reductions in prediction error and gains in rank correlation; ablations validate the contributions of propagation simulation and SMF interactions. The framework offers a scalable, dynamic, and interpretable approach to SMPP that can inform broader social media analytics and recommendation tasks.

Abstract

Accurately predicting the popularity of user-generated content (UGC) is essential for advancing social media analytics and recommendation systems. Existing approaches typically follow an inductive paradigm, where researchers train static models on historical data for popularity prediction. However, the UGC propagation is inherently a dynamic process, and static modeling based on historical features fails to capture the complex interactions and nonlinear evolution. In this paper, we propose PopSim, a novel simulation-based paradigm for social media popularity prediction (SMPP). Unlike the inductive paradigm, PopSim leverages the large language models (LLMs)-based multi-agent social network sandbox to simulate UGC propagation dynamics for popularity prediction. Specifically, to effectively model the UGC propagation process in the network, we design a social-mean-field-based agent interaction mechanism, which models the dual-channel and bidirectional individual-population interactions, enhancing agents' global perception and decision-making capabilities. In addition, we propose a multi-source information aggregation module that transforms heterogeneous social metadata into a uniform formulation for LLMs. Finally, propagation dynamics with multimodal information are fused to provide comprehensive popularity prediction. Extensive experiments on real-world datasets demonstrate that SimPop consistently outperforms the state-of-the-art methods, reducing prediction error by an average of 8.82%, offering a new perspective for research on the SMPP task.

Figures (5)

• Figure 1: Comparison of different paradigms for the SMPP task. (a) The traditional inductive SMPP paradigm constructs a static popularity prediction model by fitting historical data. (b) The proposed simulation-based paradigm dynamically simulates the UGC propagation for future popularity prediction, which achieves significant improvements.
• Figure 2: The framework of the PopSim. PopSim reformulates SMPP in a simulation-and-predict manner. In the simulation phase: (1) we construct a social network sandbox with LLM-based multi-agents to simulate social media users; (2) the agents interact through the proposed social mean field mechanism to model the UGC propagation process. The resulting social mean field state is used as UGC propagation features for popularity prediction. In the prediction phase: (3) multi-source information aggregation is developed to transform heterogeneous social media metadata and UGC propagation features into semantically unified formulations; (4) the prediction model then analyzes the above information for popularity prediction.
• Figure 3: The distribution of agent behavior and opinion scores under posts with different popularity levels.
• Figure 4: Model performance under different agent scales and interaction rounds.
• Figure 5: A case study of Popsim. The left shows the post content to be predicted, the upper-right visualizes the opinion scores of all agents in the social network, and the lower-right displays the social mean field state. The line graph represents the dynamics of the numerical mean field, the solid box indicates the textual mean field state, and the dashed box shows typical agent behavior at the current time step.