Reproduce “HyperFusion: Hierarchical Multimodal Ensemble Learning for Social Media Popularity Prediction” on OpenResearch

About this paper

Liliang Ye, Yunyao Zhang, Yafeng Wu, Yi-Ping Phoebe Chen, Junqing Yu, Wei Yang, Zikai Song

The HyperFusion framework from Huazhong University of Science and Technology and La Trobe University predicts social media post popularity using a three-tier hierarchical multimodal ensemble learning approach with a semi-supervised training strategy. It achieved third place in the SMP Challenge 2025 Image Track, demonstrating a Spearman’s Rank Correlation of 0.7324 and a Mean Absolute Error of 1.2402.

Key insights

• A hierarchical multimodal ensemble, progressively integrating features across abstraction levels, is effective for capturing complex interactions and dynamic importance of diverse data.
• User-centric attributes (e.g., follower count, past behavior) are the most influential factors in determining social media popularity, followed by semantic and visual content.
• Robust data preprocessing, including outlier filtering and cross-validation, alongside semi-supervised learning, is critical for achieving high performance and generalization in skewed social media datasets.

Summary by alphaXiv.