CasDacGCN: A Dynamic Attention-Calibrated Graph Convolutional Network for Information Popularity Prediction.

Area of Science:

• Social Network Analysis
• Information Diffusion Modeling

Background:

• Accurate prediction of information diffusion on social platforms is vital.
• Existing graph neural network methods struggle with temporal dynamics and sparse cascades.

Purpose of the Study:

• To propose a novel model, CasDacGCN, for enhanced information popularity prediction.
• To address limitations in capturing spatiotemporal features and cascade structures.

Main Methods:

• Developed Cascading Dynamic attention-calibrated Graph Convolutional Network (CasDacGCN).
• Integrated snapshot-level encoding, global temporal modeling, and cross-attention.
• Employed a hypernetwork-based sample-wise calibration strategy for adaptive representation learning.

Main Results:

• CasDacGCN demonstrated superior performance in popularity prediction.
• The model effectively fused spatiotemporal features and learned adaptive representations.
• Consistent outperformance on two real-world datasets was observed.

Conclusions:

• CasDacGCN offers an effective solution for information popularity prediction.
• The model's architecture successfully models multi-scale diffusion patterns.
• Validated effectiveness in real-world scenarios with complex cascade dynamics.