Line Graph Neural Networks for Link Prediction.

Area of Science:

• Graph theory
• Network analysis
• Machine learning

Background:

• Link prediction is a key graph analytical task with numerous applications.
• Current deep learning methods convert link prediction to graph classification, requiring graph pooling and causing information loss.

Purpose of the Study:

• To propose a novel method for graph link prediction that overcomes information loss inherent in current approaches.
• To leverage line graphs to reframe link prediction as a node classification problem.

Main Methods:

• The proposed method utilizes line graphs, where each node represents an edge in the original graph.
• Link prediction in the original graph is reformulated as a node classification task in the corresponding line graph.

Main Results:

• The novel method consistently outperforms state-of-the-art methods across fourteen diverse datasets.
• The approach demonstrates superior performance with fewer model parameters.
• High training efficiency was observed compared to existing methods.

Conclusions:

• Reframing link prediction using line graphs offers a more effective and efficient solution.
• The proposed node classification approach on line graphs mitigates information loss associated with graph pooling.