Enhancing Popularity Prediction via Bidirectional Modal Correlation Learning

Short-video Popularity Prediction (SPP) is critical in various real-world domains, including hot topic analysis and public opinion management. Existing multimodal prediction methods primarily focus on capturing the consistent relationship between modalities, often overlooking the potential antagonistic correlations. To address this challenge, we propose BMCL, a Bidirectional Modal Correlation Learning framework for short-video popularity prediction. BMCL first leverages the pretrained modality encoder to extract the valuable modality representations. Subsequently, a bidirectional modal correlation-aware module is introduced to model both positive and negative relationships between different modality-specific representations, enabling fine-grained cross-modal interaction modeling and improving the prediction accuracy. Extensive experiments on two real-world short-video datasets demonstrate that BMCL significantly outperforms six state-of-the-art baseline models.