# Towards Loosely-Coupling Knowledge Graph Embeddings and Ontology-based Reasoning

**Authors:** Zoi Kaoudi, Abelardo Carlos Martinez Lorenzo, Volker Markl  
**arXiv:** [2202.03173](https://arxiv.org/abs/2202.03173)  
**Date:** 2022-02-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.03173v2/

## Key points

- Knowledge graph completion, also known as link prediction, infers missing information from knowledge graphs
- Existing approaches relying solely on input knowledge graph often yield unsatisfactory results
- Proposed approach combines knowledge graph embeddings with domain-specific reasoning to improve accuracy
- Combination of techniques enhances prediction accuracy by incorporating domain knowledge not explicitly included in the input knowledge graph
- Users can integrate their own knowledge graph embedding and reasoning methods
- Initial results show significant improvements in mean reciprocal rank (MRR) accuracy compared to vanilla knowledge graph embeddings
- Outperforms hybrid solutions that combine knowledge graph embeddings with rule mining and reasoning by up to 3.5 times MRR
- Evaluations conducted using synthetic and real-world knowledge graphs including DBPedia and LUBM datasets
- Subset of DBPedia used due to computational limitations, consisting of a well-connected subgraph with 12 popular relations and related information
- Approach compared against several baseline models including TransE, DistMult, ComplEx, HolE, ConvE, KALE, and plogicNet
- Consistent hyperparameter settings used for both vanilla KGEs (knowledge graph embeddings) and proposed approach
- Experimentation conducted on a machine with an AMD Opteron processor running at 32x2.3 GHz, 62GB RAM, and an NVIDIA GPU.

## Summary

Knowledge graph completion, also known as link prediction, is a task that involves inferring missing information from knowledge graphs. It is widely used in various applications such as product recommendation and question answering. However, existing approaches to knowledge graph completion, which rely solely on the information present in the input knowledge graph, often yield unsatisfactory results. This limitation makes these solutions inadequate for critical domains like healthcare. To address this issue and improve the accuracy of knowledge graph completion, we propose a novel approach that combines the data-driven power of knowledge graph embeddings with domain-specific reasoning derived from experts or entailment regimes like OWL2. By loosely coupling these two techniques, we not only enhance prediction accuracy by incorporating domain knowledge that may not be explicitly included in the input knowledge graph but also allow users to integrate their own knowledge graph embedding and reasoning methods. Our initial results demonstrate significant improvements in mean reciprocal rank (MRR) accuracy compared to vanilla knowledge graph embeddings. We achieve up to 3 times enhancement in MRR accuracy and outperform hybrid solutions that combine knowledge graph embeddings with rule mining and reasoning by up to 3.5 times MRR. In order to validate our proposal, we conducted evaluations using both synthetic and real-world knowledge graphs. The datasets used include DBPedia (a real-world KG accompanied by an ontology) and LUBM (a popular synthetic KG generator). Due to computational limitations, we worked with a subset of DBPedia containing a well-connected subgraph consisting of the 12 most popular relations and related information. We compared our approach against several baseline models including TransE, DistMult, ComplEx, HolE, ConvE, KALE, and plogicNet. For each dataset and model, we utilized consistent hyperparameter settings for both vanilla KGEs (knowledge graph embeddings) and our approach. The hardware used for experimentation consisted of a machine with an AMD Opteron processor running at 32x2.3 GHz, 62GB RAM, and an NVIDIA GPU.