Abstract
Entity alignment (EA) is a critical task for integrating multi-source knowledge graphs (KGs). While Graph Convolutional Networks (GCNs) have become a standard approach, they often suffer from "neighbor noise"—uninformative or misleading connections that degrade embedding quality. This paper introduces AGEA (Adaptive Graph Convolutional Network for Entity Alignment), a lightweight, parameter-free framework designed to mitigate neighbor noise. By leveraging the Sinkhorn algorithm to dynamically calculate edge reliability and adjust weights, AGEA achieves state-of-the-art performance in both supervised and unsupervised settings with high computational efficiency.
Introduction
Knowledge Graphs (KGs) such as Wikidata, DBpedia, and YAGO provide structured representations of real-world information. However, these graphs are often siloed by language or domain, leading to fragmentation. Entity Alignment (EA) aims to identify equivalent entities across different KGs (e.g., matching "Ville de séquoia" in a French KG to "Redwood City" in an English KG).
Traditional GCN-based methods assume that similar neighbors imply entity equivalence. However, KGs frequently contain "noisy" neighbors—relations that do not contribute to alignment or provide conflicting signals. This noise can overwhelm the structural information, leading to inaccurate mathematical representations (embeddings) and poor alignment performance.
Methodology: The AGEA Framework
The AGEA framework is distinguished by its parameter-free architecture, which relies on input embeddings and structural information rather than complex trainable weight matrices. This design minimizes overfitting and enhances efficiency.
Adaptive Edge Weight Calculation
The core of AGEA is its ability to "purify" the graph structure through dynamic weighting:
- Global Similarity via Sinkhorn: Instead of local matching, AGEA employs the Sinkhorn algorithm to transform a similarity matrix into a doubly stochastic matrix. This treats alignment as a global assignment problem, ensuring a more robust initial mapping.
- Reliability Scoring: The model assesses the reliability of potential matches by calculating the margin between the primary and secondary best matches. A larger gap indicates a more confident alignment.
- Dynamic Edge Weighting: These reliability scores are used to adjust edge weights during the aggregation process. Informative edges are strengthened, while noisy edges are weakened, "purifying" the information flow.
- Relation Enhancement: The framework incorporates relationship types (e.g., "located in" vs. "born in") to further refine entity representations.
Experimental Evaluation
AGEA was evaluated on two major benchmark suites:
- DBP15K: A standard cross-lingual dataset (ZH-EN, JA-EN, FR-EN) with 15,000 aligned pairs each.
- SRPRS: A sparse dataset (EN-FR, EN-DE) designed to test performance on entities with fewer connections, reflecting real-world KG conditions where entities are often less interconnected.
Key Results and Insights
- Superior Accuracy: AGEA consistently outperformed state-of-the-art methods like RAGA and MRAEA. In supervised tests, it improved Hits@1 (the percentage of correct top-1 recommendations) by significant margins.
- Robustness to Sparsity: A notable finding is AGEA's superior performance on low-degree entities. While traditional models struggle when neighbors are few, AGEA's adaptive weighting prevents isolated noisy neighbors from distorting the entity's embedding.
- Unsupervised Power: The unsupervised variant of AGEA often surpassed the performance of previous supervised models, highlighting the strength of its structural reliability signal.
- Efficiency: Due to its parameter-free nature, AGEA is remarkably fast, completing training on DBP15K in approximately 120 seconds—significantly faster than many complex neural alternatives.
Conclusion and Future Work
AGEA addresses the neighbor noise problem in entity alignment through a lightweight, adaptive weighting mechanism. By utilizing the Sinkhorn algorithm for global assignment and reliability estimation, it provides a robust and efficient solution for KG integration. Future research will focus on relaxing the one-to-one assignment assumption to better handle "dangling entities" that exist in only one of the source graphs.
Summary updated