Who You Are and Where You Belong: Learning Individual and Group Features for Social Media Video Popularity Prediction

Abstract

Accurate popularity prediction is crucial for social media platform management and user satisfaction improvement. Video popularity is jointly influenced by the video's own content quality and the potential of the interest groups it belongs to. Prior work usually focuses on aligning and interacting multimodal features, or alternatively uses aggregated group signals. However, they overlook the collaborative contributions of individual features and group potential to popularity. Due to the complexity and diversity of a video's semantics, it often belongs to multiple groups. Moreover, samples differ in how much they rely on each source. To this end, we present an adaptive individual–group fusion framework for video popularity prediction. Specifically, it first learns content representations from the video. It then performs multi-granularity, multi-view clustering to derive comprehensive group representations. Finally, guided by the historical information of both the sample and its groups, it adaptively fuses different features to produce the popularity prediction. Our team shallow\_team achieved second place in the SMP Challenge 2026 Video Track. Extensive experiments further confirm its effectiveness. Our code is available at https://github.com/dymanne123/SMP2026.