CaSTGCN: Deep Learning Method for Information Cascade Prediction

With the advancement of online social media technologies, a vast amount of information is constantly being spreaded and interacting with users on social networks. The diffusion of online information has a significant impact on individuals, organizations, and society. Thus, controlling and predicting information cascades are particularly important. Traditional information cascades prediction and pure sequence modeling methods struggle to both capture the temporal and structural information in cascades for accurate predictions. Recently, many studies have begun to explore deep learning algorithms, particularly graph neural networks (GNNs) and sequence modeling neural network algorithms in predicting information cascades. In this context, we introduce a novel prediction model named Cascade Spatio-Temporal Graphical Convolutional Network (CaSTGCN) based on an encoder-decoder structure to capture both structural and temporal features inherent in information cascades and make prediction of their future incremental size. CaSTGCN learns the representation of cascade sequence data combined with graph data through multiple Spatial Temporal Graph Convolutional Network (ST-GCN) modules, then employs Temporal Convolutional Networks (TCN) to further explore the temporal features and make predictions. Experimental results on real-world dataset and synthetic dataset indicate that our model surpasses existing models.