Modeling Multimodal Information Cascade on Social Media with Interpretable Mixture of Experts

Information popularity prediction is a crucial yet challenging task for studying the dynamics of information diffusion patterns on social media platforms, which can benefit a wide range of applications, such as misinformation detection and viral marketing. While existing methods have achieved favorable performance in popularity predictions, they primarily focus on modeling the structural and temporal characteristics of information cascades. However, we argue in this paper that exploring other informative signals, such as textual content, can be critical to boost popularity prediction accuracy. Against this background, we propose a novel problem setting: multimodal information cascade modeling, which incorporates four essential elements, including information cascade dynamics, user profiles, textual content, and visual content, and we construct the corresponding benchmark datasets to support this problem. Subsequently, we propose MMCas, a novel approach for the MultiModal information Cascade popularity prediction task, which is designed to subtly capture the characteristics of the above four elements and, more importantly, their inherent multimodal interactions. Specifically, we first leverage diverse feature extraction pipelines of the four multimodal elements. We then design a Mixture of Experts (MoE) interaction mechanism for the modality fusion and deploy a reweighting module that assigns importance scores for the output of each interaction expert, providing both local and global interpretation. Extensive experiments conducted on * Corresponding author.