Mining user features with hyperbolic representations for diffusion prediction.

Area of Science:

• Computer Science
• Social Network Analysis
• Machine Learning

Background:

• Information diffusion prediction is crucial for real-world applications.
• Existing models struggle with social network hierarchies and propagation dynamics.
• Euclidean embeddings inadequately represent complex user relationships.

Purpose of the Study:

• To propose a novel deep learning framework, Hyper-MUF, for improved information diffusion prediction.
• To address limitations in existing models regarding network structure and temporal dynamics.
• To leverage hyperbolic representations for mining static and dynamic user features.

Main Methods:

• Reconstructing social networks in hyperbolic space to capture hierarchical relationships.
• Employing pooling and attention mechanisms to extract dynamic user features from cascade sequences.
• Integrating static and dynamic features into a unified representation for diffusion prediction.

Main Results:

• Hyper-MUF effectively models both structural hierarchy and temporal evolution in information diffusion.
• Model performance was validated on four real-world datasets.
• Prediction accuracy improved by 10% to 20% compared to state-of-the-art methods on large-scale datasets.

Conclusions:

• Hyper-MUF demonstrates superior performance in information diffusion prediction.
• The proposed hyperbolic representation and feature mining techniques are effective.
• The framework offers a unified approach to capturing complex diffusion dynamics.