Source-linked AI summary
Relational Reflection Entity Alignment
Xin Mao, Wenting Wang, Huimin Xu, Yuanbin Wu, Man Lan
TL;DR
Entity alignment must identify equivalent entities across multi-source knowledge graphs, while existing methods exhibit unexplained transformation and embedding-model weaknesses. The paper introduces Shape-Builder & Alignment and RREA, whose relational reflection transformation satisfies two derived criteria and yields consistent state-of-the-art performance, exceeding prior methods by 5.8%-10.9% on Hits@1.
Problem
Entity alignment faces unexplained failures of standard GNN linear transformations and advanced link-prediction embedding models despite increasingly complex architectures.
Method
The paper unifies alignment methods with Shape-Builder & Alignment and proposes RREA, which uses orthogonal Relational Reflection Transformation to construct relation-specific embeddings.
Results
5.8%-10.9% on Hits@1: RREA exceeds state-of-the-art methods across all real-world datasets.
Takeaways & Limitations
Relational differentiation and dimensional isometry provide transformation criteria for entity alignment, while relational reflection preserves norms and relative distances.
Takeaways & Limitations
The framework applies to alignment methods based purely on structural information and excludes methods using entity names pre-aligned by translation or cross-lingual embeddings.
Abstract
from arXiv · showhide
Entity alignment aims to identify equivalent entity pairs from different Knowledge Graphs (KGs), which is essential in integrating multi-source KGs. Recently, with the introduction of GNNs into entity alignment, the architectures of recent models have become more and more complicated. We even find two counter-intuitive phenomena within these methods: (1) The standard linear transformation in GNNs is not working well. (2) Many advanced KG embedding models designed for link prediction task perform poorly in entity alignment. In this paper, we abstract existing entity alignment methods into a unified framework, Shape-Builder & Alignment, which not only successfully explains the above phenomena but also derives two key criteria for an ideal transformation operation. Furthermore, we propose a novel GNNs-based method, Relational Reflection Entity Alignment (RREA). RREA leverages Relational Reflection Transformation to obtain relation specific embeddings for each entity in a more efficient way. The experimental results on real-world datasets show that our model significantly outperforms the state-of-the-art methods, exceeding by 5.8%-10.9% on Hits@1.
1 INTRODUCTION
Entity alignment integrates multi-source knowledge graphs by identifying equivalent entities, but increasingly complex GNN-based methods expose unexplained transformation and embedding-model failures. The paper unifies existing methods, derives transformation criteria, and proposes RREA, which outperforms state-of-the-art methods on Hits@1.
- Entity alignment identifies equivalent entity pairs across knowledge graphs, supporting knowledge integration and downstream tasks such as recommendation and search.
- Existing methods use either translation-based embeddings aligned across vector spaces or Siamese-style multi-layer GNNs trained with contrastive or triplet loss.
- Recent GNN-based architectures are increasingly complex, making the effectiveness of individual components difficult to interpret.
- Removing or weakening GNN linear transformations can reduce GCN-Align Hits@1 performance by ⩾10%, while advanced link-prediction embedding models can underperform TransE in entity alignment.
- Shape-Builder & Alignment unifies existing methods and explains that entity alignment requires orthogonal transformations to preserve norms and relative distances.
- RREA uses Relational Reflection Transformation to create relation-specific embeddings while preserving geometric structure, exceeding state-of-the-art methods by 5.8%-10.9% on Hits@1.
2 RELATED WORK
Prior entity-alignment work includes translation-based and GNN-based approaches, with GNNs introduced to capture global graph information. Relation-aware extensions seek to represent heterogeneous knowledge graphs more effectively.
- Existing entity-alignment methods are categorized by their motivations into two broad categories.
- Translation-based Methods: Translation-based methods commonly use TransE to constrain entity and relation embeddings, then align separate knowledge-graph spaces through mapping or sharing.
- GNNs-based Methods: GNN-based methods were introduced to exploit global graph information that translation-based models trained on individual triples may miss.
- GNNs-based Methods: GCN-Align uses multi-layer vanilla GCN encoders with a contrastive or triplet loss, but vanilla GCN cannot effectively utilize rich relation information in knowledge graphs.
- Relation-aware Extensions: Later models incorporate relations through concatenation, relation-dependent weights, dual relation graphs, or relation-specific transformations.
3 PRELIMINARY
Knowledge graphs represent information as entity-relation-entity triples, and entity alignment seeks matching entities across independently constructed graphs. Experiments use established cross-lingual and monolingual datasets with fixed train-test procedures.
- A knowledge graph is represented as G = (E,R,T), where E and R are entity and relation sets and T is the set of triples.
- Entity alignment finds entity pairs referring to the same real-world object across multi-source knowledge graphs.
- The alignment task uses pre-aligned seed pairs to identify additional aligned entity pairs.
- Experiments use DBP15K cross-lingual datasets and DWY100K monolingual datasets drawn from DBpedia, Wikidata, and YAGO3.
- Each DWY100K dataset contains 100,000 reference entity alignments and more than nine hundred thousand triples.
- Following prior work, 30% of aligned pairs are used for training and 70% for testing, with results averaged over five shuffled training runs.
4 A UNIFIED ENTITY ALIGNMENT FRAMEWORK
The paper unifies translation-based and GNNs-based entity alignment as Shape-Builder & Alignment, then uses this framework to explain transformation constraints and derive criteria for ideal operations.
- Shape-Builder & Alignment: Shape-Builder constrains randomly initialized embeddings into distributions with Shape Similarity, while Alignment matches those shapes using pre-aligned seeds or shared anchors.Translation-based methods fit this framework through mapping or sharing modules.
- Shape-Builder & Alignment: An unconstrained mapping can preserve seed-pair fits while destroying norms and relative distances, whereas an orthogonal matrix preserves shape similarity during transformation.Orthogonal transformation acts as a rotation operation.
- GNNs-based Methods: GNNs-based methods also fit the framework because Aggregate operations and apart losses jointly build distributions with Shape Similarity.The Aggregate operation brings similar entities closer, while apart loss separates dissimilar entities.
- Scope: The framework is outside the scope of methods that use entity names for pre-alignment, because their GNNs primarily perform noise smoothing rather than actual alignment.The stated scope is methods based purely on structural information.
- Transformation Constraints: The transformation matrix should be orthogonal because this preserves entity norms and relative distances, and experiments show this conclusion applies to simple and complex GNNs-based methods.Unit transformation is the simplest special case of an orthogonal transformation.
- Advanced KG Embeddings: Advanced KG embedding models can underperform simpler baselines in entity alignment because their transformations produce relation-specific embeddings without satisfying the alignment framework’s shape-preservation requirements.The paper reports that translation-based advanced models are at least 17% worse than TransE and GNNs-based models at least 3% worse than GCN.
5 THE PROPOSED METHOD
RREA integrates Relational Reflection Transformation into GNNs to produce relation-specific entity embeddings while preserving the geometric properties required for entity alignment.
- RREA incorporates Relational Reflection Transformation into GNNs to satisfy relational differentiation and dimensional isometry simultaneously.
- Relational Reflection Transformation: Each relation embedding defines a hyperplane, and its reflection matrix transforms entity embeddings along that relational hyperplane.
- Relational Reflection Transformation: The reflection matrix is orthogonal, preserving entity norms and relative distances during transformation.
- Relational Reflection Aggregate Layer: RREA aggregates neighboring entities using relation-specific reflection matrices and attention-like weight coefficients.
- Relational Reflection Aggregate Layer: Compared with RGCN, relational reflection uses fewer trainable parameters because each reflection matrix has d rather than d2 degrees of freedom.
- The model combines multi-hop GNN representations, dual-aspect entity-relation embeddings, triplet alignment loss, CSLS testing, and iterative semi-supervised training.
6 EXPERIMENTS
Experiments evaluate RREA under structural, semi-supervised, and textual settings, using standard entity-alignment metrics and robustness studies. RREA consistently outperforms competing methods, while ablations and low-resource tests support the effectiveness and robustness of its components.
- Evaluation Setup: Experiments use 30% of pre-aligned entity pairs for training, evaluate on the remainder, and report five-run averages using Hits@k and MRR.Hits@k measures the percentage of correctly aligned entities among the top-k candidates.
- Compared Methods: RREA provides basic, semi-supervised, and textual versions to enable comparisons across methods using structural data, generated data, and entity names.The three versions correspond to the paper’s fairness-oriented comparison categories.
- Main Results: RREA consistently ranks first across evaluation metrics, exceeding TransEdge by at least 6% on Hits@1 in the basic setting and MRAEA by more than 5% in the semi-supervised setting.RREA(semi) also improves over RREA(basic) by an average of 6% on Hits@1 through iterative generation of extra training data.
- Main Results: Textual methods show larger performance differences across datasets, while structural methods are described as more universal because textual information may be unavailable or difficult to translate reliably.The text notes that textual methods perform much better on French than on the other two languages.
- Ablation Studies: CSLS improves performance by about 4%, while the Relational Reflection Aggregate Layer and Dual-Aspect Embedding add about 15% and 7% on Hits@1, respectively.The ablation starts from GCN-Align and adds the three RREA designs incrementally.
- Robustness Analysis: RREA outperforms compared GNN methods at every tested pre-aligned ratio and exceeds 52% Hits@1 on DBP15K with only 10% of pairs used for training.This exceeds MuGNN’s performance when 40% of pre-aligned pairs are used.
7 CONCLUSIONS
The paper unifies entity-alignment methods to explain two counter-intuitive phenomena, derives transformation criteria, and proposes RREA with relational reflection. Across real-world datasets, RREA is consistently best and exceeds the state-of-the-art by more than 5.8% on Hits@1.
- The unified framework explains why standard GNN linear transformations and many advanced link-prediction embeddings perform poorly in entity alignment.
- RREA uses relational reflection to satisfy relational differentiation and dimensional isometry through a new transformation operation.
- RREA ranks consistently first across real-world datasets and outperforms the state-of-the-art by more than 5.8% on Hits@1.