Robust Classification of Information Networks by Consistent Graph Learning.

Area of Science:

• Graph-based machine learning
• Network analysis
• Data mining

Background:

• Graph regularization methods assume label-link consistency, which is often violated in real-world networks.
• Inconsistent links (connecting nodes of different classes) significantly degrade network classification performance.
• Existing methods struggle to handle the presence of these inconsistent links.

Purpose of the Study:

• To develop a novel algorithm, Consistent Graph Learning, that is robust to inconsistent links in network data.
• To improve the accuracy of network classification by learning a more consistent graph representation.
• To address the challenge posed by inconsistent links in graph regularization.

Main Methods:

• The proposed method learns a consistent network by adjusting relation matrices, reducing inconsistent links.
• It employs joint graph regularization on nuclear norm minimization of consistent relation matrices.
• It incorporates ℓ1-norm minimization on the difference matrices between original and learned relation matrices.

Main Results:

• Consistent Graph Learning demonstrates robustness against inconsistent links in network data.
• The algorithm successfully learns a network with fewer inconsistent and more consistent links.
• Experimental results show superior performance compared to state-of-the-art methods on homogeneous and heterogeneous networks.

Conclusions:

• Consistent Graph Learning effectively handles inconsistent links, a common issue in real-world networks.
• The method offers a significant improvement in network classification accuracy.
• This approach provides a robust solution for graph-based learning tasks with noisy link information.