Fusing temporal and structural information via subgraph sampling and multi-head attention for information cascade...

Area of Science:

• Computer Science
• Network Science
• Data Science

Background:

• Online social networks generate vast amounts of information cascades.
• Existing deep learning models struggle with the dynamic nature and spatiotemporal features of cascades.
• Accurate cascade prediction is crucial for research and industry applications.

Purpose of the Study:

• To develop a novel subgraph-based model, CasSubTS, for enhanced information cascade prediction.
• To address limitations in capturing dynamic changes and integrating temporal and structural information.
• To improve the accuracy and robustness of information cascade prediction methods.

Main Methods:

• CasSubTS samples time-step subgraphs to capture dynamic node changes.
• It integrates in-degree and out-degree for enhanced node representation and uses multi-head graph attention networks for multi-scale structural learning.
• Temporal information is aggregated using an attention mechanism, and channel attention prioritizes spatiotemporal feature fusion before incremental prediction via a multilayer perceptron.

Main Results:

• CasSubTS demonstrated superior performance compared to six other models on two real-world and one synthetic dataset.
• The model effectively captures dynamic cascade evolution and integrates diverse features.
• Experimental results validate the effectiveness of the proposed subgraph-based approach and attention mechanisms.

Conclusions:

• CasSubTS offers a significant advancement in information cascade prediction by effectively modeling dynamic spatiotemporal dependencies.
• The integration of subgraph sampling, multi-head attention, and channel attention enhances prediction accuracy.
• This model provides a robust framework for understanding and predicting information diffusion in online social networks.