Towards Self-Explainable Information Cascade Popularity Prediction

Cascade popularity prediction, which aims to forecast future user engagement with online content, is crucial for effective marketing and management on social media platforms. Current approaches primarily focus on modeling structural and temporal features from cascade graphs, but often overlooking noise in information propagation and lacking interpretability. To overcome these challenges, we propose ExCas, a novel graph-wise self-Explainable framework for Cascade popularity prediction. ExCas achieves accurate popularity forecasting while filtering out task-irrelevant noise, thereby enhancing model interpretability. Specifically, ExCas comprises three key components. First, the bottleneck subgraph extractor identifies influential users and constructs bottleneck subgraphs using the information bottleneck principle to capture essential structural features. Second, the propagation-aware subgraph extractor leverages distance and depth relationships among users to extract critical diffusion patterns, forming propagation-aware subgraphs. Finally, the cascade-aware graph Transformer adaptively selects the most salient characteristics from these subgraphs to generate enriched cascade representations for more accurate prediction. Extensive experiments on real-world datasets demonstrate that ExCas significantly outperforms state-of-the-art baselines.