Structural Similarity Propagation for Entity Alignment

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

Version 1 of 255 papers72 conceptsfrozen 2026-10-04

In the effort to build a global "Web of Data," researchers face a persistent bottleneck: semantic heterogeneity. When we try to link two different knowledge bases—such as an English Wikipedia graph and a Chinese Baidu Baike graph—we cannot always rely on entity names. Labels may be written in different languages, use cryptic identifiers like MA_0000436, or be entirely obfuscated.

To solve this, a long-running research direction has moved away from lexical matching toward structural similarity propagation. This approach treats the "DNA" of an entity not as its name, but as its position within a relational network.

The Principle of Structural Invariance

The core intuition behind this research is that equivalent entities across different graphs should possess structurally similar local neighborhoods. As Shanqing Yu and colleagues argue in SubGraph Networks based Entity Alignment for Cross-lingual Knowledge Graph, local network topology serves as a language-invariant feature. Even if the word for "Apple" differs between English and Chinese, the "shape" of the graph around that entity—its connections to "iPhone," "Steve Jobs," and "Cupertino"—remains remarkably consistent.

This consistency allows systems to identify matches by encoding topological "motifs" into a shared representation space. Zhichun Wang demonstrated in Cross-lingual Knowledge Graph Alignment via Graph Convolutional Networks that neighborhood-aware encoding can bridge the gap between different alphabets by mapping these local patterns into a unified vector space.

Similarity Flooding: The Recursive Logic

If we assume that similar entities have similar neighbors, we can define similarity recursively. This is the foundation of the Similarity Flooding algorithm and its derivatives. The logic follows a simple rule: if two nodes are determined to be similar, the similarity "floods" or propagates to their neighbors, increasing their likelihood of being a match.

In GMO: A Graph Matching for Ontologies, Wei Hu and Yuzhong Qu modeled ontologies as bipartite graphs. They used an iterative matrix-based process where the similarity of two entities is derived from the similarity of the statements (triples) in which they appear. This allows a small initial signal—perhaps a few "seed" matches—to spread across the entire graph.

Systems like RiMOM, developed by Juanzi Li and Jie Tang, use this propagation to resolve ambiguities. As described in RiMOM Results for OAEI 2008, structural propagation acts as a performance booster, identifying correspondences that linguistic analysis alone would miss by leveraging the hierarchical relationships of previously matched concepts.

Reaching a Fixpoint

Because the similarity of a node depends on its neighbors, and those neighbors' similarities depend on their own neighbors, the calculation is typically performed as an iterative refinement process. Gilles Bisson framed this in Learning in FOL with a Similarity Measure as a system of equations. Information flows through the relational graph, with the influence of distant entities decreasing geometrically over successive iterations.

The process continues until it reaches a fixpoint—a stable state where further iterations no longer change the similarity scores. Wei Hu noted that these scores typically stabilize after about twelve iterations, suggesting that global structural information propagates efficiently through the network.

From Classical Matching to Neural Embeddings

Modern research has bridged the gap between these classical iterative algorithms and deep learning. Zequn Sun and Jiacheng Huang proved in What Makes Entities Similar? A Similarity Flooding Perspective for Multi-sourced Knowledge Graph Embeddings that many popular embedding-based alignment models are actually seeking a mathematical fixpoint of structural similarity. In essence, Graph Convolutional Networks (GCNs) are a modern, high-dimensional implementation of the similarity flooding principle.

This perspective has led to more robust frameworks like PipEA. In Understanding and Guiding Weakly Supervised Entity Alignment with Potential Isomorphism Propagation, Yuanyi Wang and colleagues use a propagation operator to bridge the "information gap" in settings where labeled data is scarce. By treating alignment as a similarity propagation task rather than just feature extraction, they can refine the outputs of various encoders without intensive retraining.

Why It Works: The Power of Context

The success of this direction across decades—from Falcon-AO in 2006 to Lily in 2020—stems from a fundamental truth about information: context is identity.

Whether aligning cross-lingual text streams as Tao Ge did in Fine-grained Coordinated Cross-lingual Text Stream Alignment, or mapping relational databases to ontologies in IncMap, the structural approach succeeds because it treats entities as nodes in a web of relationships. By propagating similarity through these links, researchers can recover the underlying semantics of a dataset even when its surface-level labels are missing, translated, or intentionally obscured.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 58 papers
  2. Version 1 · 2026-10-04 · 55 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_structural_similarity_propagation_7d4ee50b,
  title = {structural similarity propagation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structural-similarity-propagation/rdr_7d4ee50b1f39461692f46be6182a897b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_structural_similarity_propagation_for_en_7d4ee50b_v1,
  title = {Structural Similarity Propagation for Entity Alignment},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structural-similarity-propagation-for-entity-alignment/rdr_7d4ee50b1f39461692f46be6182a897b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}