# Knowledge Enhanced Graph Neural Networks

**Authors:** Luisa Werner, Nabil Layaïda, Pierre Genevès, Sarah Chlyah  
**arXiv:** [2303.15487](https://arxiv.org/abs/2303.15487)  
**Date:** 2023-03-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.15487v1/

## Key points

- Graph data is ubiquitous across various fields such as natural science, social networks, and semantic web.
- Graph completion tasks like node classification and link prediction are challenging due to incomplete and noisy graphs.
- Researchers have explored the use of neural methods like graph neural networks (GNNs) for learning rich representations of noisy graphs.
- Symbolic methods enable exact reasoning on graphs.
- The authors propose a neuro-symbolic framework called KeGNN that combines both paradigms to allow for the integration of prior knowledge into a GNN model.
- KeGNN comprises a GNN as a base on which knowledge enhancement layers are stacked with the objective of refining predictions with respect to prior knowledge.
- KeGNN is evaluated on multiple benchmark datasets for node classification and demonstrates its effectiveness in improving performance compared to existing state-of-the-art models.
- KeGNN presents a promising approach towards leveraging both neural and symbolic methods for learning on graph data while incorporating prior knowledge to improve accuracy.

## Summary

The ubiquity of graph data across various fields such as natural science, social networks, and semantic web has led to an increased interest in graph completion tasks like node classification and link prediction. However, graphs are often incomplete and noisy, making it challenging to extract meaningful information from them. To address this issue, researchers have explored the use of neural methods like graph neural networks (GNNs) for learning rich representations of noisy graphs. At the same time, symbolic methods enable exact reasoning on graphs. In this paper titled "Knowledge Enhanced Graph Neural Networks," the authors propose a neuro-symbolic framework called KeGNN that combines both paradigms to allow for the integration of prior knowledge into a GNN model. KeGNN comprises a GNN as a base on which knowledge enhancement layers are stacked with the objective of refining predictions with respect to prior knowledge. The authors instantiate KeGNN in conjunction with two standard GNNs: Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs). They evaluate KeGNN on multiple benchmark datasets for node classification and demonstrate its effectiveness in improving performance compared to existing state-of-the-art models. Overall, KeGNN presents a promising approach towards leveraging both neural and symbolic methods for learning on graph data while incorporating prior knowledge to improve accuracy. This work has significant implications for various applications that rely on accurate predictions from incomplete or noisy graphs.