Source-linked AI summary
Neighborhood Matching Network for Entity Alignment
Yuting Wu, Xiao Liu, Yansong Feng, Zheng Wang, Dongyan Zhao
TL;DR
Entity alignment is challenged by incomplete, structurally heterogeneous knowledge graphs whose equivalent entities can have dissimilar neighborhoods. NMN samples informative neighbors and jointly matches neighborhood differences across graphs, achieving the best and more robust performance across datasets and evaluation metrics against 12 recent methods. Its sampling strategy can be unstable when alignment clues are very sparse.
Problem
Entity alignment must handle equivalent entities with dissimilar neighborhood structures caused by incomplete and heterogeneous knowledge graphs.
Method
NMN combines GCN-based topology modeling, informative-neighbor sampling, and cross-graph attention matching of sampled neighborhoods.
Results
NMN achieves the best and more robust performance, consistently outperforming 12 competitive embedding-based methods across datasets and evaluation metrics.
Takeaways & Limitations
Integrating neighbor connection and relevance information enables NMN to align entities in real-world knowledge graphs with neighborhood heterogeneity.
Takeaways & Limitations
When alignment clues are very sparse, NMN’s relatively simple sampling strategy may be unstable and insufficiently robust.
Abstract
from arXiv · showhide
Structural heterogeneity between knowledge graphs is an outstanding challenge for entity alignment. This paper presents Neighborhood Matching Network (NMN), a novel entity alignment framework for tackling the structural heterogeneity challenge. NMN estimates the similarities between entities to capture both the topological structure and the neighborhood difference. It provides two innovative components for better learning representations for entity alignment. It first uses a novel graph sampling method to distill a discriminative neighborhood for each entity. It then adopts a cross-graph neighborhood matching module to jointly encode the neighborhood difference for a given entity pair. Such strategies allow NMN to effectively construct matching-oriented entity representations while ignoring noisy neighbors that have a negative impact on the alignment task. Extensive experiments performed on three entity alignment datasets show that NMN can well estimate the neighborhood similarity in more tough cases and significantly outperforms 12 previous state-of-the-art methods.
1 Introduction
Entity alignment supports knowledge integration but is difficult because real-world knowledge graphs are incomplete and structurally heterogeneous. NMN addresses this by selecting informative neighbors and matching neighborhoods across graphs, achieving robust performance on challenging benchmarks.
- Motivation: NMN targets neighborhood heterogeneity, where equivalent entities may have different neighborhood sizes, topology, or nondiscriminative common neighbors.Such heterogeneity is widespread: over 86% of equivalent pairs in DBP15K have different neighbors, reaching 90% in some language pairs.
- Approach: NMN samples each entity’s most informative neighbors to construct a discriminative neighborhood for alignment.The graph sampling strategy is designed to retain neighbors that are useful for entity alignment while reducing noise.
- Approach: NMN uses cross-graph attention-based matching to jointly compare two entities’ discriminative neighborhoods and encode their differences.The matching module complements GCN-based modeling of topological connections by modeling neighbor relevance across graphs.
2 Related Work
Prior entity-alignment work relies mainly on embedding-based representations, GNNs, and graph-matching techniques. NMN combines informative-neighbor sampling with cross-graph neighborhood matching to address the limitations of treating heterogeneous or noisy neighborhoods indiscriminately.
- Embedding-based entity alignment: Embedding-based methods map equivalent entities into a unified vector space using KG structure and seed alignments.Alignment is performed by measuring distances between learned entity embeddings.
- Embedding-based entity alignment: Prior methods often ignore neighborhood heterogeneity, while MuGNN requires pre-aligned entities and relations and AliNet treats all one-hop neighbors as equally important.These design choices can increase labeling overhead or introduce noise from uninformative neighbors.
- Graph neural networks: GNNs learn node representations by recursively aggregating neighboring-node representations, providing a structural basis for entity-alignment models.The related work surveys GCN and other GNN variants used to model KG structure.
- Graph matching and sampling: Inspired by graph matching networks, NMN applies cross-graph attention-based matching to compare sampled neighborhood subgraphs.Its graph-sampling component selects informative neighbors and constructs a neighborhood subgraph for each entity.
3 Our Approach
NMN aligns entities by combining GCN-based structural embeddings with informative neighborhood sampling, cross-graph neighbor matching, and neighborhood aggregation. This pipeline produces matching-oriented entity representations that account for heterogeneous neighborhoods and noisy neighbors.
- NMN represents each knowledge graph as entities, relations, and triples, then seeks equivalent entity pairs between two graphs using pre-aligned seeds.
- NMN follows four stages: KG structure embedding, neighborhood sampling, neighborhood matching, and neighborhood aggregation for final alignment embeddings.
- 3.2 KG Structure Embedding: GCNs aggregate higher-degree neighborhood structure, while highway networks control accumulated noise propagation across layers.NMN initializes GCN inputs with pretrained entity-name word embeddings.
- 3.3 Neighborhood Sampling: NMN samples one-hop neighbors according to contextual relatedness, constructing a discriminative neighborhood subgraph that excludes less informative neighbors from later processing.Neighbors more contextually related to the central entity are more likely to be sampled.
- 3.4 Neighborhood Matching: For each candidate pair, cross-graph attention matches neighbors across sampled subgraphs and computes vectors representing differences from the closest counterpart neighbors.Similar neighbor representations yield near-zero matching vectors, whereas differing representations are amplified through propagation.
- 3.5 Neighborhood Aggregation: NMN aggregates sampled-neighbor representations with central-entity GCN outputs to construct the matching-oriented embeddings used for alignment.Candidate selection first limits matching to entities sampled from the counterpart graph’s embedding-nearest candidates.
4 Experimental Setup
NMN is evaluated on established and newly constructed entity-alignment datasets under standard train/test settings, against 12 prior methods and controlled model variants. Performance is measured with Hits@1 and Hits@10, alongside implementation configurations for each dataset.
- Datasets: NMN is evaluated on DBP15K and DWY100K using a 30% training and 70% testing split, plus sparse S-DBP15K variants with increased cross-graph neighborhood differences.S-DBP15K is created by randomly removing triples from the non-English knowledge graph.
- Comparison models: The comparison includes 12 recently proposed embedding-based entity-alignment methods, with RDGCN and GMNN additionally using entity names.
- Model variants: Two variants isolate NMN components by removing neighborhood matching or neighborhood sampling from the full model.The first averages sampled neighbor representations, whereas the second matches all one-hop neighbors.
- Implementation details: NMN uses dataset-specific sampling settings, with five neighbors on DBP15K and DWY100K and 3 or 10 neighbors across S-DBP15K language pairs.The model uses a 2-layer GCN, 300-dimensional hidden representations, 50-dimensional neighborhood representations, and a candidate set of 20 entities.
- Implementation details: Node features are initialized from translated English entity names for DBP15K and directly from pretrained word vectors for DWY100K.
- Metrics: Evaluation uses Hits@1 and Hits@10, defined as the proportions of correctly aligned entities ranked within the top 1 or top 10 results.Higher Hits@k scores indicate better performance.
5 Experimental Results
NMN consistently outperforms competing entity-alignment methods by combining informative neighborhood sampling with cross-graph neighborhood matching. Experiments show these components are especially valuable under severe neighborhood heterogeneity and sparse alignment clues.
- 5.1 Performance on DBP15K and DWY100K: NMN substantially outperforms all alternative approaches across DBP15K and DWY100K, with its full implementation accurately modeling entity neighborhoods through sampling and matching.Table 3 covers performance across both datasets.
- 5.1 Performance on DBP15K and DWY100K: Removing neighborhood matching reduces average Hits@1 by 2.5% and Hits@10 by 0.6%, confirming its role in identifying matching neighbors and estimating neighborhood similarity.The Hits@1 difference reaches 3.9% on DBP15KFR−EN.
- 5.2 Performance on S-DBP15K: NMN outperforms BootEA, GMNN, and RDGCN on all sparse S-DBP15K datasets, demonstrating effectiveness and robustness when alignment clues are limited.Entity-name semantics improve baseline performance over structure-only methods, while NMN performs best among the compared models.
- 5.2 Performance on S-DBP15K: On DBP15KZH−EN, NMN achieves Hits@1 of 65%, 53%, and 48% when equivalent entities differ by more than 10, 20, and 30 neighbors, respectively.These results exceed RDGCN, the best-performing baseline, in increasingly heterogeneous cases.
- 5.3 Analysis: NMN’s designed sampling strategy consistently beats random sampling, while larger sampling sizes can degrade performance on the sparsest S-DBP15KZH−EN dataset.Both methods plateau at a sampling size of 3 on S-DBP15KZH−EN, after which performance declines.
6 Conclusion
NMN addresses neighborhood heterogeneity in knowledge graphs through sampling-based neighbor selection and joint estimation of topology and neighborhood similarity. Across real-world datasets, it consistently outperforms 12 recent embedding-based methods.
- NMN selects the most informative neighbors for each entity using a new sampling-based approach.
- NMN consistently outperforms 12 recent embedding-based methods across datasets and evaluation metrics.The reported performance is also described as the best and more robust among the compared methods.
- NMN estimates entity similarity from both topological structure and neighborhood similarity to address neighborhood heterogeneity.