CausE: Towards Causal Knowledge Graph Embedding

TL;DR

CausE introduces a causality-based framework for knowledge graph embedding that improves the stability and accuracy of link prediction by addressing confounders through causal intervention and disentanglement.

Contribution

This paper presents a novel causality-enhanced KGE framework, CausE, which incorporates causal intervention and disentanglement to improve knowledge graph completion.

Findings

CausE outperforms baseline models in KGC tasks.

CausE achieves state-of-the-art performance on benchmark datasets.

CausE demonstrates robustness against noisy links and trivial patterns.

Abstract

Knowledge graph embedding (KGE) focuses on representing the entities and relations of a knowledge graph (KG) into the continuous vector spaces, which can be employed to predict the missing triples to achieve knowledge graph completion (KGC). However, KGE models often only briefly learn structural correlations of triple data and embeddings would be misled by the trivial patterns and noisy links in real-world KGs. To address this issue, we build the new paradigm of KGE in the context of causality and embedding disentanglement. We further propose a Causality-enhanced knowledge graph Embedding (CausE) framework. CausE employs causal intervention to estimate the causal effect of the confounder embeddings and design new training objectives to make stable predictions. Experimental results demonstrate that CausE could outperform the baseline models and achieve state-of-the-art KGC performance. We release our code in https://github.com/zjukg/CausE.