Predicting Social Media Popularity With Large Language Models: Transforming Metadata Into Semantic-Enriched and Contextualized Text

Accurate prediction of social media content popularity remains a significant challenge due to the limitations of traditional machine learning and deep learning approaches. These methods often rely on manually engineered features that fail to capture the full complexity and nuances of social media data. A pervasive oversight is the neglect of the crucial relationship between field names (labels or identifiers) and their corresponding field values within social media data structures. Disregarding this relationship leads to issues such as loss of context, incomplete feature representation, difficulty in feature engineering, and lack of domain knowledge integration. To address these limitations, we propose a novel approach that transforms social media metadata into semantic-enriched and contextualized text representations tailored for the context of social media posts. This semantic enrichment process expands field names and values into coherent textual descriptions, effectively conveying implicit meanings and linking disparate metadata elements. We leverage these enriched semantic representations to adapt large pre-trained language models using Low-Rank Adaptation (LoRA). By integrating LoRA, the language models gain proficiency in interpreting and learning from the transformed semantic-rich social media text, facilitating a deeper understanding compared to conventional metadata-based approaches. Our principal innovation lies in substantially elevating the precision of social media popularity predictions by incorporating comprehensive semantic data descriptions into the modeling process. The LoRA-based approach enables efficient task-specific optimization while preserving the extensive knowledge base of the pre-trained models. Extensive experiments on the SMPD dataset demonstrate state-of-the-art performance, validating our method’s theoretical novelty and practical effectiveness.