Abstract
Knowledge Graph (KG) completion through ontology reasoning is a critical task for data integration. While Graph Neural Networks (GNNs) offer powerful learning capabilities, they often suffer from limited receptive fields and a lack of inherent logical reasoning. This paper presents 2-Hop GAT, an extension of Graph Attention Networks that incorporates two-hop connectivity to expand the receptive field without increasing network depth. Furthermore, we introduce Filtered 2-Hop GAT, which leverages specific RDF entailment rules—subclass transitivity and type-subclass relations—to guide the attention mechanism. Evaluated on the CaLiGraph and ORE datasets, our neuro-symbolic approach demonstrates significant improvements in reasoning performance, particularly in class subsumption tasks, achieving a Hits@10 score of 0.803 on CaLiGraph10e5.
Introduction
Knowledge Graphs and ontologies are foundational to modern data systems but are frequently incomplete. Reasoning—the process of inferring new facts from existing data—is essential to address these gaps. While Graph Neural Networks (GNNs) have emerged as a dominant paradigm for graph learning, they face two significant challenges in the context of ontology reasoning:
- Limited Receptive Field: Standard GNN layers only aggregate information from immediate (one-hop) neighbors. Increasing the receptive field typically requires stacking multiple layers, which often leads to "over-smoothing," where node representations become indistinguishable.
- Lack of Logical Grounding: GNNs are primarily data-driven and do not inherently respect formal logical rules, such as the transitivity of subclass relations.
This work proposes a scalable, neuro-symbolic framework that explicitly integrates logical RDF rules into the GNN architecture to enhance reasoning over large-scale ontologies.
Methodology
The proposed approach introduces two architectural variants designed to bridge the gap between neural learning and symbolic reasoning.
2-Hop GAT
The 2-Hop GAT model extends the standard Graph Attention Network by allowing each node to attend to both its immediate neighbors and its two-hop neighbors. By calculating node embeddings through the aggregation of messages from a wider neighborhood, the model captures longer-range dependencies without the need for deep architectures. This effectively increases the "kernel" width rather than the network depth, mitigating over-smoothing.
Filtered 2-Hop GAT
To improve efficiency and reduce noise, the Filtered 2-Hop GAT introduces a neuro-symbolic filtering mechanism. Instead of attending to all possible two-hop paths, the model focuses exclusively on paths that satisfy specific RDF entailment rules:
- Subclass Transitivity: If $A$ is a subclass of $B$, and $B$ is a subclass of $C$, then $A$ is a subclass of $C$.
- Type-Subclass Rule: If $A$ is of type $B$, and $B$ is a subclass of $C$, then $A$ is of type $C$.
By guiding the attention mechanism with these rules, the model prioritizes logically relevant information, significantly improving reasoning accuracy on complex ontologies.
Experimental Evaluation
The models were evaluated on datasets from the SemRec Challenge at ISWC 2023, including OWL2Bench, ORE (OWL Reasoner Evaluation), and CaLiGraph.
Key Results
The experimental results highlight the efficacy of expanding the receptive field:
- CaLiGraph10e5 Performance: The Filtered 2-Hop GAT achieved a Hits@10 of 0.803 for class subsumption. In contrast, the standard GAT achieved only 0.006, demonstrating that standard architectures struggle to learn transitivity without explicit guidance.
- Scalability and Efficiency: By using a direct message-passing schema instead of computationally expensive random walks (e.g.,
metapath2vec), the 2-Hop GAT maintains high efficiency even on large datasets. - Dataset Sensitivity: The "Filtered" approach is most effective on datasets with a significant proportion of subclass relations, such as ORE and CaLiGraph, where logical transitivity plays a major role.
Conclusion and Future Work
The Enhanced GAT framework demonstrates that incorporating two-hop connectivity and symbolic RDF rules significantly improves the reasoning capabilities of GNNs on large-scale ontologies. Future research will focus on integrating complex logical axioms from the T-box (schema) and expanding the set of supported OWL rules to handle more sophisticated logical inferences.
Summary updated