Self-Training Enhanced: Network Embedding and Overlapping Community Detection With Adversarial Learning.

Area of Science:

• Graph theory
• Machine learning
• Network science

Background:

• Network embedding (NE) encodes vertex relationships into low-dimensional spaces for tasks like vertex classification.
• Existing NE models often rely on local connectivity, neglecting multifaceted vertex roles in overlapping communities.
• Real-world networks exhibit complex structures where vertices can belong to multiple, non-exclusive communities.

Purpose of the Study:

• To propose an adversarial learning approach (ACNE) for modeling overlapping communities in networks.
• To enhance ACNE with a perception-based walking strategy and a self-training mechanism (ACNE-ST).
• To improve the performance of network embedding on vertex classification and overlapping community detection tasks.

Main Methods:

• Developed ACNE, an adversarial learning framework mapping community-vertex associations into an embedding space.
• Implemented a perception-based walking strategy during initialization to identify boundary vertices.
• Integrated a self-training mechanism (ACNE-ST) using soft community assignments for weight updates.

Main Results:

• ACNE effectively models overlapping communities by embedding community-vertex associations.
• The ACNE-ST variant demonstrated improved performance through its self-training approach.
• Both ACNE and ACNE-ST outperformed state-of-the-art methods in vertex classification and overlapping community detection.

Conclusions:

• The proposed ACNE and ACNE-ST methods offer effective solutions for network embedding with overlapping communities.
• ACNE advances network representation learning by capturing multifaceted vertex roles.
• The enhanced methods show significant improvements in downstream network analysis tasks.