DeepRicci: Self-supervised Graph Structure-Feature Co-Refinement for Alleviating Over-squashing

The paper presents DeepRicci, a self-supervised model for graph structure-feature co-refinement. It aims to solve the over-squashing issue that is common in Graph Neural Networks (GNNs). This is achieved by using the concept of Ricci curvature from Riemannian geometry. The model refines node features through geometric contrastive learning and refines graph structure using backward Ricci flow. The experiments show that DeepRicci outperforms other methods and it effectively addresses the over-squashing issue. The code for the model is available on GitHub.

Publication date: 24 Jan 2024
Project Page: https://github.com/RiemanGraph/
Paper: https://arxiv.org/pdf/2401.12780