Abstract
Entity alignment (EA) is a fundamental task for integrating knowledge graphs (KGs) from disparate sources. Traditional Graph Neural Network (GNN) approaches often falter due to "neighborhood heterogeneity," where the same real-world entity is described by different sets of neighbors across KGs. This paper introduces RRDGNN (Relational Reflective Disentangled Graph Neural Network), a framework designed to learn disentangled, multi-aspect representations of entities. By utilizing relational reflective routing and alignment-aware self-attention, RRDGNN effectively bridges the gap between heterogeneous neighborhoods. Our evaluation on standard benchmarks shows that RRDGNN consistently outperforms existing state-of-the-art models, especially in sparse data environments.
Introduction: The Neighborhood Heterogeneity Problem
Knowledge Graphs (KGs) serve as the backbone for various AI applications, yet they are frequently siloed. Entity Alignment (EA) aims to identify equivalent entities across these KGs. While GNNs have become the standard for this task, they typically aggregate all neighboring information into a single static embedding. This leads to the "neighborhood heterogeneity" problem: if one KG focuses on an entity's personal life while another focuses on their professional career, a standard GNN may fail to recognize them as the same entity.
As shown in Figures 1 and 2, the neighborhoods for "Stephen Curry" in two different KGs can be almost entirely disjoint. RRDGNN addresses this by "disentangling" these various aspects—such as family, career, and location—allowing the model to focus on the specific facets that overlap.
Methodology
RRDGNN introduces several key innovations to the GNN architecture:
1. Micro-disentanglement via Relational Reflective Routing
Instead of a single representation, RRDGNN learns $K$ independent embeddings for each entity. It employs a "relational reflective" mechanism where the semantics of the relations (edges) guide the clustering of neighbors into specific aspects. For instance, neighbors linked by "father" or "brother" are routed to a "family" aspect embedding.
2. Macro-disentanglement and Regularization
To prevent the $K$ aspects from becoming redundant, the model incorporates Mutual Information (MI) regularization. This constraint forces the learned aspects to remain independent and diverse, ensuring that each captures a unique "facet" of the entity.
3. Alignment-aware Self-attention
During the alignment phase, the model uses a self-attention strategy to adaptively weight the different aspects. If two entities show a strong match in their "career" aspect but differ in "family" (perhaps due to missing data), the model learns to prioritize the career aspect for that specific alignment pair.
4. Multi-aspect Seed Generators
In semi-supervised settings, RRDGNN iteratively expands its training set with new "seed" alignments. It utilizes multiple generators tailored to different neighborhood densities, ensuring high accuracy for both well-connected and "long-tail" (sparse) entities.
Experimental Evaluation
The model was rigorously tested on two major benchmark suites: DBP15K (multilingual) and S-DBP15K (sparse).
Key Results
- Performance: RRDGNN achieved a Hits@1 of 0.851 on the DBP15K ZH-EN subset, a 3% improvement over the previous state-of-the-art (SPSEA).
- Robustness: In tests where 50% of triples were removed to simulate extreme heterogeneity, RRDGNN maintained a significant lead (6.2%) over competitive baselines like Dual-AMN.
- Ablation Study: Removing the disentanglement component led to a performance drop of 3.4% to 4.4% in Hits@1, confirming that multi-aspect learning is critical for handling heterogeneous KGs.
Conclusion
RRDGNN effectively mitigates the neighborhood heterogeneity problem by learning disentangled entity representations guided by relational semantics. Its ability to adaptively weight different aspects makes it highly robust to noise and data sparsity. Future work will explore the application of this disentanglement framework to other graph-based tasks such as Knowledge Graph Question Answering (KGQA) and recommendation systems.
Summary updated