New Benchmark for Social Media Popularity Prediction.

Predicting the future reach or influence of online content from its early stages is a critical task in social media, enabling applications like advertising optimization and strategic content planning. However, existing research often fails to integrate multimodal content (text, visual) with temporal social interaction signals, and the field is fragmented across various datasets, modalities, and evaluation protocols. To overcome these challenges, this paper introduces MMG-Pop, a comprehensive Multi-modal Graph-based Popularity Prediction benchmark. MMG-Pop unifies datasets from platforms like Bluesky and Reddit, incorporates diverse modalities and temporal interaction signals, and standardizes evaluation protocols, facilitating fair comparisons. Alongside the benchmark, the researchers propose MMG-PopNet, a unified multi-modal graph-based network specifically designed to jointly model these heterogeneous signals and graph-structured social interactions. Extensive experiments on MMG-Pop demonstrate MMG-PopNet's superior performance compared to representative baselines. The study also yields valuable insights into cross-platform training generalization, the benefits of multi-task prediction, the contributions of different modalities, and the limitations of LLM-based prediction. These findings lay a robust foundation for future research in modeling social dynamics and developing socially-aware agentic ecosystems.