Chien-Chun Ni, Yu-Yao Lin, Feng Luo, Jie Gao (2019), Scientific Reports 9, 9984 (2019) — ARXIV-1907.03993
Key takeaways
- Iterating an Ollivier-Ricci-curvature-driven flow on edge weights stretches inter-community (negatively curved) edges and shrinks intra-community (positively curved) ones, so thresholding the evolved weights (“surgery”) recovers communities with accuracy competitive with or better than standard methods on SBM, LFR and several labelled real networks.
Seeded from the abstract and a skim, before the reading in NOTE-006. What follows is what the work says about itself.
Real networks often contain communities, and most detection methods are statistical or combinatorial. The authors instead treat a network as a geometric object and communities as a geometric decomposition, borrowing curvature and discrete Ricci flow, tools that decompose smooth manifolds in mathematics. They apply this to networks with known ground-truth communities and report that the geometric approach is effective.
Standing in the record
Filed by the reading-time triage of 2026-09-25: 770 seconds of active reading over 5 sessions in the papers-feed tracker. It was filed Deferred, unread. NOTE-006 is the close reading of 2026-09-25, and it placed the work: Active — the source paper for Ollivier-Ricci-flow community detection and the GraphRicciCurvature package, and worth reading for the method (§3, SI §C). It should be cited for the algorithm, not for superiority. Its SBM figure shows it losing to Spinglass and Fast Greedy past p_inter/p_intra = 0.5. Its theorem is weaker than stated, and it covers a parameter setting (α = p = 0) that differs from the one used in experiments (α = ½, p = 2).
It is also held in the Anthology of the SOTA, as ANTH-LIT-766, filed on 2026-10-01 under the anthology’s new graphs-and-networks topic during its issue 180. Holding it in both follows ADR-013: the anthology reads it beside its graph-learning material, and this record for network science.
When seeded, before the reading in NOTE-006, the entry gave a priority for a deeper reading: low — the algorithm and headline results are fully captured by this skim and the owner spent ~13 min (t=770 s); only read deeper if implementing Ricci-flow clustering or needing the SI proof.
What the seeding asked a deeper reading to check (NOTE-006 is that reading):
- The empirical superiority claims rest on a limited benchmark set (SBM, LFR, GNet, ~6 labelled real graphs) compared against iGraph baselines; check sensitivity to the surgery threshold and number of iterations, which are hand-set.
- Theorem 4.1 covers only a highly symmetric toy family; whether any guarantee extends to SBM-like graphs is not shown here.
- ML link: Ollivier-Ricci curvature and the GraphRicciCurvature tooling are used in later graph-learning work (e.g. curvature-based analyses of GNN message passing); this paper is the community-detection source, not a GNN paper.
Access when seeded: arXiv abs page and full v1 PDF (29 pp. incl. supplementary information) read via arxiv.org. DOI confirmed via Crossref (Sci. Rep. 9, article 9984, published 2019-07-10). The paper lists no keywords; those above are chosen from its text. Supplementary proofs (Theorem 4.1) not checked.