Robust Knowledge Graph Embedding via Denoising

TL;DR

This work addresses the vulnerability of knowledge graph embeddings to embedding-space perturbations and proposes Robust Knowledge Graph Embedding via Denoising (RKGE-D). By reframing KGEs as energy-based models and tying denoising to score matching, RKGE-D learns robust representations while a randomized smoothing–based certification scheme provides formal robustness guarantees via the certified radius (CR) and metrics like ACR and CA. Empirical results on FB15k-237 show RKGE-D outperforms state-of-the-art baselines under noisy embeddings, with enhanced resilience in link prediction and multi-hop reasoning tasks. The approach offers a principled path toward reliable KG reasoning in noisier real-world settings, while highlighting sensitivity to hyperparameters and suggesting future work on adaptive noise strategies.

Abstract

We focus on obtaining robust knowledge graph embedding under perturbation in the embedding space. To address these challenges, we introduce a novel framework, Robust Knowledge Graph Embedding via Denoising, which enhances the robustness of KGE models on noisy triples. By treating KGE methods as energy-based models, we leverage the established connection between denoising and score matching, enabling the training of a robust denoising KGE model. Furthermore, we propose certified robustness evaluation metrics for KGE methods based on the concept of randomized smoothing. Through comprehensive experiments on benchmark datasets, our framework consistently shows superior performance compared to existing state-of-the-art KGE methods when faced with perturbed entity embedding.

Figures (3)

• Figure 1: Link prediction shift caused by embedding level perturbation.
• Figure 2: Hyperparameter sensitivity
• Figure 3: Case Study

Theorems & Definitions (1)

• Definition 1: $CR$ in link prediction