Predicting information popularity in social networks has become a central focus of network analysis. While recent advancements have been made, most existing approaches rely solely on the final cascade size as the primary supervision signal for model optimization. This narrow focus limits the model generalization ability, particularly when faced with highly heterogeneous cascades. Additionally, in real-world scenarios, obtaining detailed social relationships is challenging, complicating effective structural feature learning. To address these issues, this paper proposes a semi-supervised model called Dual Variational Cascade AutoEncoders (DVCAE), which leverages parallel structural and temporal variational autoencoders for enhanced feature learning and popularity prediction. The model first aggregates multiple cascades into a global interaction graph, enabling structural information sharing across cascades. Then, it applies sparse matrix factorization-based graph embedding and graph filtering techniques on global and local cascade graphs respectively, generating initial node embeddings that are insensitive to topological perturbations. After that, two parallel variational autoencoders are designed to generate hidden representations for structural and temporal features respectively, with two self-supervised reconstruction losses integrated into the prediction loss to enrich supervision signals. Extensive experiments conducted on three real-world datasets demonstrate that DVCAE outperforms state-of-the-art models in terms of prediction accuracy.

Information popularity prediction on social media platforms is a valuable and challenging issue.However, existing studies either neglect the correlation among different cascades, or lack a comprehensive consideration of user behavioral proximity and preference with respect to different messages.In this paper we propose a graph neural network-based framework named HeDAN (heterogeneous diffusion attention network), which comprehensively considers various factors affecting the information diffusion to predict the information popularity more accurately.Specifically, we first construct a heterogeneous diffusion graph with two types of nodes (user and message) and three types of relations (Friendship, Interaction, and Interest).Among them, Friendship reflects the strength of social relationship between users, Interaction reflects the behavioral proximity between users, and Interest reflects user preference to messages.Next, a graph neural network model with hierarchical attention mechanism is proposed to learn from these relations.Specifically, at the nodelevel, we utilize the graph attention network to learn the subgraph structure and generate the representations of nodes under each specific relationship.At the semantic-level, we distinguish the importance of different nodes in different relations via multihead self-attention mechanism.Extensive experimental results on three datasets show the superior performance of our proposed model over the state-of-the-arts.