Abstract
Entity Alignment (EA) is critical for integrating knowledge from disparate Knowledge Graphs (KGs). While Graph Neural Networks (GNNs) currently dominate this field, they suffer from significant scalability and interpretability issues. We present LightEA, a non-neural framework that replaces complex GNNs with an efficient "Three-view Label Propagation" algorithm. By leveraging random orthogonal label generation and sparse Sinkhorn iterations, LightEA achieves state-of-the-art accuracy while being up to 10x faster than deep learning counterparts and capable of scaling to graphs with over one million entities.
Introduction
Knowledge Graphs (KGs) such as Wikidata and DBpedia store vast amounts of information as (head, relation, tail) triples. Because these graphs are often developed independently, they frequently contain overlapping information. Entity Alignment (EA) is the process of identifying equivalent entities across different graphs (e.g., matching "Apple Inc." in one graph to "Apple" in another).
Currently, over 90% of EA methods rely on Graph Neural Networks (GNNs). Despite their power, GNNs possess two major "inborn defects":
- Weak Scalability: Real-world KGs contain millions of entities. The iterative forward and backward propagation required by GNNs is computationally expensive and memory-intensive.
- Poor Interpretability: GNNs act as "black boxes," making it difficult to explain why specific entities were aligned—a significant drawback for high-stakes data integration tasks.
Methodology
LightEA abandons neural architectures in favor of a reinvented Label Propagation (LP) algorithm designed for heterogeneous graphs. The framework consists of three core innovations:
1. Random Orthogonal Label Generation
Standard LP uses "one-hot" vectors, which do not scale to millions of entities. LightEA utilizes high-dimensional random unit vectors. Due to the mathematical "curse of dimensionality," these random vectors are nearly orthogonal, allowing them to represent distinct entities efficiently without the massive memory footprint of one-hot matrices.
2. Three-view Label Propagation
KGs are complex 3D structures (head, relation, tail). To avoid the computational nightmare of 3D tensor products, LightEA projects the 3D adjacency tensor into three 2D views: Top (Relation-Tail), Side (Head-Tail), and Front (Head-Relation). This allows labels to propagate across different connection types efficiently.
3. Sparse Sinkhorn Iteration
EA is essentially a 1-to-1 assignment problem. While the Sinkhorn algorithm is the standard tool for this, its $O(N^2)$ complexity is prohibitive. LightEA employs a sparse version that only considers the top-$k$ nearest neighbors, reducing complexity to $O(NK)$.
Experimental Results
LightEA was evaluated on several benchmarks, including DBP15K, DWY100K, and the massive DBP1M (1 million entities).
Performance and Scalability
- Speed: LightEA is remarkably fast. On the DBP15K dataset, it runs in just 7 seconds. On the 100K dataset, it finishes in under 35 seconds—roughly 1/10th the time of previous state-of-the-art GNN models.
- Scalability: LightEA successfully processed the DBP1M dataset, whereas many GNN-based methods fail due to GPU memory limits at this scale.
- Accuracy: Despite its lightweight nature, LightEA-L (the version incorporating entity name features) outperformed or matched complex models like Dual-AMN and SEU across all benchmarks.
Key Insights and Conclusion
LightEA demonstrates that for graph alignment, the structural propagation of information is often more critical than complex non-linear feature transformations. By using a linear approach, the model becomes inherently interpretable; researchers can "trace" the propagation to see exactly which neighbors and relations contributed to an alignment score.
While future work remains in optimizing hardware efficiency and multi-GPU support, LightEA proves that non-neural, linear frameworks can provide a scalable, robust, and transparent alternative to GNNs for large-scale Knowledge Graph integration.
Summary updated