Improving Network Representation Learning via Dynamic Random Walk, Self-Attention and Vertex Attributes-Driven...

Area of Science:

• Network Science
• Machine Learning
• Data Mining

Background:

• Network representation learning (NRL) methods, including random walk and neural language models, are vital for analyzing complex interactions.
• Existing NRL approaches struggle with preserving local community structures and integrating diverse vertex attributes due to fixed random walks and limited feature extraction capabilities.

Purpose of the Study:

• To propose a general NRL framework that addresses the deficiencies of existing methods.
• To enhance the analysis of network data by effectively capturing both structural and attribute information.

Main Methods:

• Developed an asymmetric similarity and h-hop dynamic random walk strategy to preserve local community structures.
• Employed a self-attention-based sequence prediction model for learning local and global structural features.
• Introduced an attributes-driven Laplacian space optimization for integrating structural and attribute features.

Main Results:

• The proposed dynamic structure and vertex attribute fusion network embedding framework demonstrated superior performance.
• Evaluations on benchmark datasets using node visualization and classification confirmed the effectiveness of the approach.
• The method successfully preserved local community structure and integrated vertex attributes.

Conclusions:

• The novel NRL framework offers a significant advancement in network data analysis.
• This approach provides a more robust and versatile method for understanding complex network interactions.
• The findings suggest broader applicability in various network-related research areas.