nucleation

THEORY-1159: Aligning three networks jointly, through one multi-marginal fused Gromov–Wasserstein coupling with an anchor-based random-walk cost and a cluster decomposition, finds held-out correspondences better than pairwise and existing multi-network methods in most cases, as the paper's plots show, in both pairwise and whole-tuple hits, on benchmarks whose true correspondences are consistent across the three networks by construction; whether the joint coupling itself is the cause is not separated from the new cost and the clustering

2 min read

Properties5
StatusActive
Version2
Tagsnetwork-science, mathematics
SourceLIT-1814 — Hierarchical Multi-Marginal Optimal Transport for Network Alignment
Cited as “extends” byTHEORY-1210 — When several networks are aligned, fitting one joint multi-marginal coupling finds held-out correspondences better than fitting them in pairs, and the gain comes from the joint fit itself, not from the cost or the search space that come with it

Graph View

Source

Zeng et al. (2024), LIT-1814, §5 (pp. 5–7), Appendices C–D (pp. 12–15) of arXiv 2310.04470v2; read in NOTE-1334.

What was actually shown

Design. Three networks per benchmark: ER-500 (each a noisy permutation of one base graph, p. 15), Douban-230, DBLP-500/1000 and, with 17 attributes, ACM(A) and DBLP(A) (Table 4, p. 14). Ten random splits, 10% of the true node triples as anchors and 90% for testing (p. 5). Thirteen baselines: consistency-, embedding- and OT-based; the pairwise ones are combined into a tensor by multiplying their three alignment matrices (p. 5). Metrics: pairwise Hits@K, high-order Hits@K (the whole triple in the top K) and MRR (p. 5).

Result. HOT is reported best “in most cases” (p. 5), by “at least 12.0% on plain networks in terms of high-order Hits@10” (p. 1), with the larger margins on plain networks and on high-order metrics (p. 5). Read off the plots, roughly: on ER-500 pairwise Hits@K near 99% and high-order near 80%. It also wins on pairwise hits, so on these data the joint fit did not cost pairwise accuracy. Results for larger DBLP graphs (Tables 1–2) have no baseline.

What could have failed. The joint fit could have lost pairwise accuracy to its consistency constraint, or a pairwise method with a good cost could have matched it. Neither happened against these baselines.

What this does not say

  • That joint fitting is the cause. HOT introduces three things at once: the RWR cost from shared anchors, the cluster decomposition, and the joint coupling. The paper attributes the high-order gains to joint alignment (p. 5) without an ablation. On attributed DBLP the plain version does nearly as well, which it credits to the cost (p. 13).
  • That consistency never loses local distinctions. Every benchmark’s truth is a consistent set of triples, so a consistency constraint cannot be wrong there. The record’s interest (CLAIM-496, QUESTION-319) is in data where it can be.
  • Anything about independently fitted pairwise alignments composed and compared with direct ones. The baselines’ product rule never composes maps.
  • Exact values. The main results are plots; numbers above are readings.

Note of 2026-10-30: narrowed and promoted

The title now keeps the paper’s “in most cases” (p. 5). The paper’s attribution of the gain to joint fitting, with the ablation that would test it, is THEORY-1210, Proposed, which extends this entry.

Backlinks