D ² : Customizing Two-Stage Graph Neural Networks for Early Rumor Detection through Cascade Diffusion Prediction

Early rumor detection is crucial for mitigating the widespread dissemination of misinformation. Existing methods predominantly rely on complete rumor diffusion graphs, which are challenging to obtain in real-world scenarios, complicating early detection efforts. To address this challenge, we propose D2, a two-stage framework for early rumor Detection, integrating cascade Diffusion prediction. This framework aims to enhance early rumor detection by incorporating diffusion prediction capabilities. Specifically, a dynamic heterogeneous graph neural network (GNN) is developed to jointly model users' social and propagation graphs, enabling accurate prediction of potential diffusion paths using limited observed data within short time windows. The inferred diffusion paths are then integrated with early-stage data, and GNNs are employed for graph classification. However, the varying data distributions across different social media platforms necessitate extensive tuning to optimize GNN architectures. To facilitate the detection of rumor diffusion graphs at the initial stages, a search space is designed across four dimensions- aggregation, merge, readout, and sequence functions-encompassing various GNN architectures. Subsequently, D2 employs an efficient differentiable search algorithm to identify high-performance GNNs within this search space. Experimental results on real social media datasets demonstrate that this approach significantly improves both the accuracy and robustness of early rumor detection.