Abstract
Entity Alignment (EA) aims to find equivalent entity pairs between Knowledge Graphs (KGs), which is the core step of bridging and integrating multi-source KGs. In this paper, we argue that existing GNN-based EA methods inherit the inborn defects from their neural network lineage: weak scalability and poor interpretability. Inspired by recent studies, we reinvent the Label Propagation algorithm to effectively run on KGs and propose a non-neural EA framework -- LightEA, consisting of three efficient components: (i) Random Orthogonal Label Generation, (ii) Three-view Label Propagation, and (iii) Sparse Sinkhorn Iteration. According to the extensive experiments on public datasets, LightEA has impressive scalability, robustness, and interpretability. With a mere tenth of time consumption, LightEA achieves comparable results to state-of-the-art methods across all datasets and even surpasses them on many.
1. Introduction
Entity Alignment (EA) is a critical task in knowledge engineering, focused on identifying identical entities across disparate Knowledge Graphs (KGs). As KGs like Wikidata, DBpedia, and YAGO grow to encompass millions of entities, the ability to integrate them efficiently becomes paramount.
In recent years, Graph Neural Networks (GNNs) have become the de-facto standard for EA. By learning low-dimensional embeddings that capture structural information, GNNs have pushed the boundaries of alignment accuracy. However, this progress comes at a cost. GNN-based methods are often computationally expensive, difficult to scale to million-node graphs, and lack the transparency required for high-stakes data integration. This paper introduces LightEA, a non-neural framework that returns to the fundamentals of Label Propagation (LP) to achieve superior efficiency and interpretability without sacrificing performance.
2. The Case Against GNN Dominance
While GNNs are powerful, they suffer from three primary "inborn defects" in the context of EA:
- Weak Scalability: The iterative nature of forward and backward propagation in deep networks requires significant GPU memory and time. Training on large-scale KGs can take days.
- Poor Interpretability: GNNs function as "black boxes," making it nearly impossible to trace why a specific alignment was predicted.
- Diminishing Returns: Recent research has focused on adding complex mechanisms—such as relational attention or hyperbolic embeddings—which increase model complexity for marginal gains in accuracy.
3. The LightEA Framework
LightEA moves away from the neural paradigm entirely, reinventing the classical Label Propagation algorithm for the complex, heterogeneous structure of KGs.
3.1 Random Orthogonal Label Generation
Standard LP uses one-hot vectors to represent labels, which is memory-intensive ($O(N^2)$) for large graphs. LightEA instead uses independent random unit vectors on a high-dimensional hyper-sphere. Due to the properties of high-dimensional geometry, these vectors are approximately orthogonal. This allows the model to maintain the performance of one-hot encoding while drastically reducing the memory footprint.
3.2 Three-view Label Propagation
KGs consist of triples (head, relation, tail), forming a 3D adjacency tensor. Standard LP only works on homogeneous 2D graphs. LightEA solves this by "projecting" the 3D tensor into three 2D views:
- Side View: Head-to-tail transitions.
- Front View: Head-to-relation associations.
- Top View: Relation-to-tail associations. Labels are propagated across these views, capturing the interplay between entities and relations without the cubic complexity of full tensor products.
3.3 Sparse Sinkhorn Iteration
EA requires a one-to-one mapping between entities. While the Sinkhorn algorithm is the standard way to enforce this constraint, it is computationally expensive ($O(N^2)$). LightEA introduces a "Sparse" version that only considers the top-$k$ nearest neighbors for each entity (retrieved via FAISS), bringing the complexity down to $O(Nk)$.
4. Experimental Evaluation
LightEA was evaluated on several established benchmarks, including DBP15K, SRPRS, DWY100K, and the million-node DBP1M dataset.
4.1 Efficiency and Scalability
LightEA is remarkably fast. It obtains results on DBP15K in just 7 seconds and handles DWY100K in under 35 seconds. This is roughly 1/10th the time required by state-of-the-art GNN methods like Dual-AMN. Furthermore, while many GNN-based methods cannot run on the DBP1M dataset due to memory constraints, LightEA handles it easily on a single GPU.
4.2 Accuracy and Robustness
Despite being non-neural, LightEA-I (the iterative version) and LightEA-L (which incorporates literal features like entity names) consistently match or outperform complex GNN models across all benchmarks. This suggests that the core benefit of GNNs in EA is "feature smoothing"—an effect that LP achieves more efficiently.
4.3 Interpretability
Because the model is linear, the authors can "trace" the propagation. Unlike neural embeddings, LightEA allows researchers to see exactly which neighboring entities and relations contributed to a correct (or incorrect) alignment at each step, providing a clear path for error analysis.
5. Conclusion
LightEA demonstrates that complex neural architectures are not always necessary for high-performance Entity Alignment. By reinventing Label Propagation with random orthogonal vectors and a three-view projection strategy, LightEA offers a scalable, robust, and interpretable alternative to GNNs. Future work will focus on refactoring the implementation into pure CUDA for even greater speedups and testing on multi-GPU clusters.
Summary updated