Deep Network Embedding for Graph Representation Learning in Signed Networks.

Area of Science:

• Graph mining and network analysis
• Machine learning and deep learning

Background:

• Network embedding learns low-dimensional node representations for graph mining.
• Existing methods primarily focus on unsigned networks, neglecting distinct properties of signed networks with positive and negative links.

Purpose of the Study:

• To propose a deep network embedding model for signed networks that preserves structural balance.
• To improve graph representation learning for signed networks.

Main Methods:

• A semisupervised stacked auto-encoder reconstructs adjacency connections in signed networks.
• A larger penalty is imposed on reconstructing scarce negative links compared to abundant positive links.
• Pairwise constraints ensure positively connected nodes are closer than negatively connected nodes in the embedding space.

Main Results:

• The proposed model learns effective low-dimensional node vector representations for signed networks.
• The model demonstrates superiority over state-of-the-art algorithms in link sign prediction and community detection tasks.
• Experimental results on real-world datasets validate the model's effectiveness.

Conclusions:

• The developed deep network embedding model successfully addresses the challenges of signed network representation learning.
• The approach effectively preserves structural balance, leading to improved performance in downstream tasks.
• This work offers a significant advancement in applying network embedding to signed graph data.