LargeEA: Aligning Entities for Large-scale Knowledge Graphs

Congcong Ge, Xiaoze Liu, Lu Chen, Baihua Zheng, Yunjun Gao

Abstract

Entity Alignment (EA) is a fundamental task for integrating multi-source Knowledge Graphs (KGs). While existing Graph Neural Network (GNN) based methods have shown promise on small-scale benchmarks, they frequently encounter scalability bottlenecks—such as out-of-memory (OOM) errors—when applied to real-world KGs containing millions of entities. This paper introduces LargeEA, a scalable framework that decomposes the alignment task into collaborative Structure and Name channels. By utilizing a novel collaborative partitioning strategy and name-feature fusion, LargeEA achieves state-of-the-art performance on million-node graphs while significantly reducing computational overhead.

The Scalability Challenge in Entity Alignment

Knowledge Graphs like DBpedia and YAGO are essential for semantic search and recommendation engines, but they are often siloed and incomplete. Entity Alignment (EA) seeks to identify equivalent entities across these disparate graphs to merge them into a unified resource.

Most current EA methods rely on GNNs, which require loading the entire graph structure into memory. While effective for benchmarks with 15,000 to 100,000 entities, these models typically fail on million-scale graphs. There is a critical need for a framework that can handle massive KGs on standard hardware without sacrificing alignment accuracy.

The LargeEA Framework

LargeEA addresses the scalability problem through a dual-channel architecture that separates structural learning from linguistic features.

The LargeEA framework architecture featuring Name and Structure channels.

1. The Structure Channel

To avoid memory exhaustion, LargeEA partitions KGs into manageable mini-batches. The core innovation is METIS-CPS (Collaborative Partition Strategy). Unlike standard random partitioning, METIS-CPS re-weights edges to ensure that known equivalent entities (seeds) and their immediate neighborhoods are grouped into the same mini-batch. This "entity locality" allows GNN models to be trained independently on each batch while preserving the structural context necessary for accurate alignment.

2. The Name Channel

Recognizing that entity names provide a strong alignment signal, the Name Feature Fusion (NFF) module combines semantic similarity (via BERT embeddings) with string similarity (via Levenshtein distance). Furthermore, LargeEA employs name-based data augmentation to generate "pseudo-seeds." By identifying entities with high mutual name similarity across graphs, the model can bootstrap its training even when human-labeled data is limited.

Experimental Evaluation

The researchers evaluated LargeEA on standard cross-lingual benchmarks (IDS15K, IDS100K) and a newly introduced large-scale benchmark, DBP1M, which contains approximately one million entities per KG.

Performance comparison (Hits@1) across different dataset scales for the EN-FR alignment task.

Key Results

  • Breaking the Scale Barrier: LargeEA successfully aligned the DBP1M dataset, a task where competing models failed due to memory limitations.
  • Efficiency Gains: On the IDS100K dataset, LargeEA was 10x faster than the previous leading model, BERT-INT, and required 9x less GPU memory.
  • Superior Accuracy: LargeEA consistently outperformed existing models across all metrics. For instance, on the IDS100K English-French dataset, it achieved a Hits@1 of ~84%, significantly higher than standard GNN-based competitors.

Conclusion and Future Work

LargeEA demonstrates that smart partitioning combined with robust name-feature integration can overcome the scalability barriers of traditional EA methods. While the framework relies heavily on naming conventions, it sets a new benchmark for million-scale entity alignment. Future research will explore methods that maintain this scalability in environments where entity names are uninformative or missing, potentially leveraging other attributes like images or numerical data.

Summary updated

BibTeX
@article{ge2021largeea,
  author = {Congcong Ge and Xiaoze Liu and Lu Chen and Baihua Zheng and Yunjun Gao},
  title = {{LargeEA: Aligning Entities for Large-scale Knowledge Graphs}},
  journal = {Proc. VLDB Endow.},
  year = {2021},
  doi = {10.14778/3489496.3489504},
}
Figure