Toward Noise-Resistant Graph Embedding With Subspace Clustering Information.

Area of Science:

• Graph embedding
• Network science
• Machine learning

Background:

• Existing attributed network embedding methods often rely on the homophily assumption, which fails with noisy or irrelevant node attributes.
• Real-world networks frequently contain attributes that do not align with the homophily principle, limiting embedding performance.

Purpose of the Study:

• To propose a noise-resistant graph embedding method (NGE) that overcomes limitations of the homophily assumption.
• To effectively integrate network topology and attribute information despite attribute noise.
• To enhance performance in community detection, link prediction, and node classification tasks.

Main Methods:

• Constructing a tensor representation of the attributed network.
• Mapping the tensor into feature subspaces using tensor decomposition to capture community structures.
• Incorporating link-level and community-level constraints for structure embedding.
• Utilizing a feature-selection constraint for attribute embedding to manage noisy attributes.

Main Results:

• NGE demonstrates superior performance compared to state-of-the-art methods.
• The method effectively leverages subspace clustering information for robust network embedding.
• Experimental results validate the effectiveness of NGE in handling noisy attributes.

Conclusions:

• NGE provides a robust solution for attributed network embedding in the presence of noisy attributes.
• The proposed subspace clustering approach enhances the accuracy of community detection, link prediction, and node classification.
• NGE offers a significant advancement in analyzing complex, real-world networks with attribute noise.