Learning From Negative Links.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Network Analysis

Background:

• Graph convolutional networks (GCNs) excel in semisupervised node classification by aggregating neighborhood information.
• Existing GCN methods primarily utilize positive links, assuming feature similarity between connected nodes.
• The limitations of solely relying on positive links can hinder comprehensive node representation.

Purpose of the Study:

• To develop a novel GCN-based learning framework that enhances node representation inference.
• To improve classification performance by incorporating negative links into graph topologies.
• To adaptively generate and integrate negative links that represent inverse correlations between nodes.

Main Methods:

• A novel GCN learning framework incorporating negative links was developed.
• Negative links were adaptively generated using a neural-network-based generation model.
• The negative link generation model was jointly optimized with the GCN for class inference.

Main Results:

• The proposed framework demonstrated improved node representation inference capabilities.
• Experimental results showed the framework achieved better or matched performance against state-of-the-art methods.
• The inclusion of adaptively generated negative links proved beneficial for classification performance.

Conclusions:

• The novel GCN framework effectively leverages negative links for enhanced node classification.
• Jointly optimizing negative link generation and GCNs leads to superior performance.
• This approach offers a promising direction for advancing graph-based semisupervised learning.