With the rapid growth of global social media users, platforms have become crucial venues for shaping public opinions. However, most existing multimodal topic models focus on large-scale content analysis and often fail to capture nuanced patterns in individual user content. To address this gap, we propose Social Media Account Topic Modeling (SMATM), a novel account-level multimodal topic modeling framework designed to enhance the analysis of individual users’ multimodal data and extract fine-grained personalized themes. SMATM innovatively enhances the ability of topic models to capture social media account data characteristics by introducing a label weighting module, employs a flexible parameter learning module to improve the model’s adaptability to users’ sparse content, and proposes novel evaluation metrics that leverage large language models’ (LLMs) understanding of complex contexts to enhance model interpretability. Experiments conducted on a multimodal account-level dataset collected from social media marketing scenarios validated the effectiveness of the SMATM model in account-level topic extraction. Compared with existing baseline models and ablation models, SMATM achieved significant leads in both topic consistency and diversity evaluation metrics. Particularly in terms of interpretability, the SMATM framework elevated theme accuracy from a baseline of 0.1864 to 0.6059. This 3.25-fold enhancement underscores the model’s superior capability for analyzing individual user behavior and multimodal data. Visualization of cross-industry account themes further confirms that SMATM produces semantically consistent and relevant topics by effectively integrating textual and visual information. Overall, the SMATM model represents a substantial advance in social media content analysis, user profiling, and personalized recommendations.
Song et al. (Tue,) studied this question.