Graph Autoencoder with Preserving Node Attribute Similarity.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Unsupervised Learning

Background:

• Graph autoencoders (GAEs) are effective for unsupervised learning on graph data.
• Existing GAEs often prioritize reconstructing graph topology over node attributes, limiting representation quality.
• Failure to preserve node attribute information weakens GAEs' learning capabilities.

Purpose of the Study:

• To propose a novel GAE model that effectively preserves node attribute similarity.
• To enhance the learning of node attributes within the GAE framework.
• To improve the overall quality of learned graph representations.

Main Methods:

• Integrated structural and attribute neighbor graphs using a novel fusion strategy.
• Employed a shared encoder to aggregate node attributes from structural and attribute neighborhoods.
• Utilized dual decoders to reconstruct both the adjacency matrix and node attribute similarity matrix.

Main Results:

• The proposed GAE model successfully preserves both structural and node attribute similarity.
• Achieved superior performance compared to state-of-the-art algorithms.
• Demonstrated significant improvements in link prediction and node clustering tasks on citation networks.

Conclusions:

• The novel GAE approach effectively fuses structural and attribute information for richer node representations.
• Preserving node attribute similarity is crucial for enhancing GAE performance.
• The method offers a promising advancement for unsupervised graph representation learning.