Piecewise-linear Ricci curvature flows on weighted graphs

TL;DR

This paper introduces a unified framework for piecewise-linear Ricci curvature flows on weighted graphs, proving their global existence, uniqueness, and effectiveness in community detection across real-world datasets.

Contribution

It establishes a unified theoretical framework for Ricci curvature flows on graphs, proving their properties and demonstrating their practical utility in community detection.

Findings

Flows have global existence and uniqueness.

Homogeneous Ricci curvatures lead to constant curvature after surgeries.

Flow outperforms baseline models on real-world datasets.

Abstract

Community detection is an important problem in graph neural networks. Recently, algorithms based on Ricci curvature flows have gained significant attention. It was suggested by Ollivier (2009), and applied to community detection by Ni et al (2019) and Lai et al (2022). Its mathematical theory was due to Bai et al (2024) and Li-M\"unch (2025). In particular, solutions to some of these flows have existence, uniqueness and convergence. However, a unified theoretical framework has not yet been established in this field. In the current study, we propose several unified piecewise-linear Ricci curvature flows with respect to arbitrarily selected Ricci curvatures. First, we prove that the flows have global existence and uniqueness. Second, we show that if the Ricci curvature being used is homogeneous, then after undergoing multiple surgeries, the evolving graph has a constant Ricci curvature on each connected component. Note that five commonly used Ricci curvatures, which were respectively defined by Ollivier, Lin-Lu-Yau, Forman, Menger and Haantjes, are all homogeneous, and that the proof of all these results is independent of the choice of the specific Ricci curvature. Third, as an application, we apply the discrete piecewise-linear Ricci curvature flow with surgeries to the problem of community detection. On three real-world datasets, the flow consistently outperforms baseline models and existing methods. Complementary experiments on synthetic graphs further confirm its scalability and robustness. Compared with existing algorithms, our algorithm has two advantages: it does not require curvature calculations at each iteration, and the iterative process converges.