Robust Knowledge Graph Embedding via Denoising

Knowledge graph embedding models have achieved remarkable success in link prediction and reasoning tasks, yet they remain highly vulnerable to perturbations in the embedding space. Such perturbations, whether introduced by noisy triples, representation drift or adversarial manipulation, can lead to severe degradation in prediction stability and significantly affect downstream multi-hop reasoning processes. To address this challenge, we propose a unified robustness enhancement framework named Robust Knowledge Graph Embedding via Denoising. The framework explicitly incorporates denoising as an auxiliary learning signal and views knowledge graph embedding models as energy-based systems, allowing us to exploit the theoretical connection between denoising objectives and score matching. This enables the model to learn stable gradients with respect to perturbed representations and improves resilience against embedding-level noise. In addition, we introduce certified robustness metrics for knowledge graph embedding based on randomized smoothing, offering a principled way to measure the certified radius within which model predictions remain unchanged. Extensive experiments on widely used benchmark datasets demonstrate that the proposed framework consistently improves both predictive performance and robustness across various categories of knowledge graph embedding models. The results further show that our method is effective under substantial perturbations and offers meaningful gains in multi-hop reasoning scenarios, highlighting its potential as a general robustness enhancement strategy for knowledge graph representation learning. Our code is available at https://github.com/tewiSong/RKGE.