# CausE: Towards Causal Knowledge Graph Embedding

**Authors:** Yichi Zhang, Wen Zhang  
**arXiv:** [2307.11610](https://arxiv.org/abs/2307.11610)  
**Date:** 2023-07-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.11610v1/

## Key points

- Knowledge graph embedding (KGE) represents entities and relations in a knowledge graph as continuous vector spaces
- KGE can be used for predicting missing triples and achieving knowledge graph completion (KGC)
- The Causality-enhanced knowledge graph Embedding (CausE) framework is proposed to address limitations of traditional KGE models
- CausE utilizes causal intervention to estimate the causal effect of confounder embeddings and introduces new training objectives for stable predictions
- Experimental results show that CausE outperforms baseline models and achieves state-of-the-art performance in KGC tasks, with a 1.4% improvement in Hit@1 on the FB15K-237 benchmark dataset
- CausE consistently outperforms other models across various metrics, as summarized in Table 1
- CausE is a universal framework applicable to different KGE models, effectively handling noisy data and improving the plausibility of predicted triples
- CausE introduces an innovative approach for knowledge graph embedding with a focus on causality and embedding disentanglement
- The code for implementing CausE is available on GitHub for further exploration and application.

## Summary

Knowledge graph embedding (KGE) is a technique that aims to represent entities and relations in a knowledge graph (KG) as continuous vector spaces. This representation can be used to predict missing triples and achieve knowledge graph completion (KGC). To address the issue of traditional KGE models only capturing brief structural correlations in triple data, leading to embeddings that are influenced by trivial patterns and noisy links present in real-world KGs, the authors of this paper propose a new paradigm for KGE: the Causality-enhanced knowledge graph Embedding (CausE) framework. It utilizes causal intervention to estimate the causal effect of confounder embeddings and designs new training objectives to ensure stable predictions. Experimental results demonstrate that CausE outperforms baseline models and achieves state-of-the-art performance in KGC tasks. Specifically, it shows a 1.4% improvement in Hit@1 on the FB15K-237 benchmark dataset. The results are summarized in Table 1, where CausE consistently outperforms other models across various metrics. Furthermore, the authors emphasize that CausE is a universal framework that can be applied to different KGE models. It effectively addresses the challenge of handling noisy data and improves the plausibility of predicted triples. In conclusion, this paper introduces CausE as an innovative approach for knowledge graph embedding with a focus on causality and embedding disentanglement. The experimental results highlight its superiority over existing methods and establish it as a new state-of-the-art technique for knowledge graph completion. The code for implementing CausE is made available on GitHub for further exploration and application.