Network Representation Learning With Community Awareness and Its Applications in Brain Networks

Area of Science:

• Network Science
• Machine Learning
• Data Mining

Background:

• Existing network representation learning methods primarily focus on microscopic (pairwise) relationships.
• The mesoscopic structure, specifically community structure, remains underexplored in network representation learning.
• Real-world networks possess essential mesoscopic properties that influence their overall behavior and function.

Purpose of the Study:

• To propose a novel deep attributed network representation learning framework with community awareness (DANRL-CA).
• To effectively capture both microscopic and mesoscopic network structures for improved representation learning.
• To develop variants that integrate community information and attribute semantics for enhanced network analysis.

Main Methods:

• Designed a neighborhood enhancement autoencoder module to capture 2-step node pair relations.
• Constructed a community-aware skip-gram module to explore multi-step relations.
• Introduced two variants, DANRL-CA-AM and DANRL-CA-CSM, differing in how community and attribute information are integrated.

Main Results:

• Evaluated DANRL-CA variants against state-of-the-art methods on four datasets for node classification and link prediction.
• Demonstrated the scalability and effectiveness of the proposed methods on various networks, including a brain network.
• DANRL-CA-CSM showed superiority, especially on networks with sparse topology and attributes, by flexibly coordinating attribute and community information.

Conclusions:

• The proposed DANRL-CA framework effectively incorporates community structure into network representation learning.
• DANRL-CA variants show significant improvements in node classification and link prediction tasks.
• DANRL-CA-CSM offers a more flexible and superior approach for networks with sparse characteristics.