High-Precision Strategies for Transitive Entity Resolution

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 216 papers32 conceptsfrozen 2026-10-04

In entity resolution and coreference tasks, we aim to determine which digital records—whether they are mentions in a text, nodes in a knowledge graph, or advertisements on the web—refer to the same real-world entity. At first glance, this seems like a simple pairwise classification problem: is Record A the same as Record B?

However, identity is a transitive relationship. If $A=B$ and $B=C$, then $A=C$ must also be true. This logical property is the "glue" that allows us to build large clusters of information, but it is also a source of extreme fragility. Across various domains, researchers have found that a single false positive link can act as a bridge, merging two massive, unrelated networks into a single meaningless "supercluster."

The Transitivity Trap and Cluster Collapse

The fundamental challenge in large-scale entity resolution is that errors do not stay local. In Resurrecting Address Clustering in Bitcoin, Malte Möser describes this as cluster collapse. In the context of blockchain analysis, a single incorrect heuristic that links two addresses can propagate through the entire transaction graph, destroying the integrity of the dataset for forensic use.

This "snowball effect" is particularly dangerous as datasets grow. Matt Barnes and colleagues in Performance Bounds for Pairwise Entity Resolution demonstrate that precision often collapses as dataset size increases. Because the number of potential false matches grows quadratically with the number of records, the cost of a single mistake scales disproportionately. As Chirag Nagpal and Kyle Miller observe in their work on Human Trafficking Online, a single false positive link can merge distinct trafficking networks, rendering the resulting data uninterpretable for law enforcement.

Local Inference vs Global Consistency

To combat this, researchers often choose between "local" and "global" inference protocols.

  • Global protocols (like All-Link) attempt to find a globally consistent partition where all items in a cluster are linked.
  • Local protocols (like Best-Link) focus on finding the single most confident connection for each mention.

In Inference Protocols for Coreference Resolution, Alla Rozovskaya, Dan Roth, and their team found that simpler, local inference often outperforms complex global clustering. Global methods are highly sensitive to "noise" in pairwise scores; a single high-scoring error can force a cascade of incorrect transitive decisions. By focusing on the "Best-Link," the system limits the opportunity for error propagation.

Similarly, in the Semantic Web, the owl:sameAs predicate is often misused to link "similar" rather than "identical" things. Gerard de Melo and Joe Raad (in The sameAs Problem) highlight how these "near-miss" links trigger chain reactions that corrupt decentralized knowledge bases.

Why Precision-First Strategies Work

The consensus across these papers—from Author Name Disambiguation by Zhenyu Zhang to Scholarly Metadata Cleaning by Athar Sefid—is a strict prioritization of precision over recall.

This preference is rooted in a practical asymmetry: it is much easier to merge two clusters later than it is to split a giant, collapsed cluster. Marek Śmieja and colleagues, in their work on Semi-Supervised Clustering ($S^3C^2$), argue that because "must-link" constraints are transitive, we should only apply them when confidence is extremely high.

The Graph Intuition: Connectivity without Density

The most compelling intuition for why high-precision, low-recall linking works comes from graph theory. In End-to-End Face Detection and Cast Grouping, SouYoung Jin uses Erdös-Rényi graph principles to show that you don't need to find every link between related records to form a correct cluster.

A cluster of $N$ records only needs a small number of valid internal links (a "spanning tree") to become a single connected component. By setting an extremely high threshold for establishing a link:

  1. You drastically reduce the probability of a false positive link (protecting precision).
  2. You still maintain high cluster integrity (recall) because the underlying graph structure naturally bridges the gaps between records.

This "sparse but certain" approach allows systems to resolve millions of entities with near-perfect precision, ensuring that the resulting data structures remain distinct and meaningful rather than collapsing into a single, uninformative aggregate.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 16 papers
  2. Version 1 · 2026-10-04 · 16 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_transitive_error_propagation_0080e04e,
  title = {transitive error propagation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/transitive-error-propagation/rdr_0080e04ebd4d4de7acde1b4e4c15816a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_precision_strategies_for_transitive_0080e04e_v1,
  title = {High-Precision Strategies for Transitive Entity Resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-precision-strategies-for-transitive-entity-resolution/rdr_0080e04ebd4d4de7acde1b4e4c15816a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}