Effective and lightweight representation learning for signed bipartite graphs.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Signed bipartite graphs model complex relationships in e-commerce and peer review.
• Existing Graph Neural Network (GNN) methods struggle with over-smoothing and inefficiency.
• Current approaches often add edges, increasing complexity and vulnerability to noise.

Purpose of the Study:

• To propose ELISE, a lightweight GNN-based method for effective node representation learning in signed bipartite graphs.
• To address limitations of existing methods, including over-smoothing, noise sensitivity, and inefficiency.
• To enhance the accuracy and efficiency of learning node embeddings in signed bipartite networks.

Main Methods:

• Extended personalized propagation to signed bipartite graphs, integrating edge signs without adding new edges.
• Employed joint learning of node embeddings on a low-rank graph approximation.
• Developed a lightweight GNN architecture named ELISE.

Main Results:

• ELISE effectively mitigates over-smoothing by incorporating signed edges directly.
• Low-rank approximation reduces noise and enhances expressiveness without compromising efficiency.
• ELISE demonstrated superior performance in link sign prediction on real-world datasets compared to existing methods.

Conclusions:

• ELISE offers an efficient and effective solution for node representation learning in signed bipartite graphs.
• The method achieves faster training and inference speeds.
• ELISE provides a robust alternative to existing GNNs for signed bipartite network analysis.